How to import and export models in PuLP
Exporting a model can be useful when the building time takes too long or when the model needs to be passed to another computer to solve. Or any other reason. PuLP offers two ways to export a model: to an mps file or to a dictionary /json file. Each offers advantages over the other.
The mps format is an industry standard. But it is not very flexible so some information cannot be stored. It stores only variables and constraints. It does not store the values of variables.
The dictionary/ json format is made to fit how pulp stores the information and so it does not lose information: this format file saves enough data to be able to restore a complete pulp model on reading it.
The interface to import and export for both formats is similar as can be seen in the Example 1 below.
Considerations
The following considerations need to be taken into account:
Variable names need to be unique. PuLP permits having variable names because it uses an internal code for each one. But we do not export that code. So we identify variables by their name only.
Variables are not exported in a grouped way. This means that if you have several dictionaries of many variables each you will end up with a very long list of variables. This can be seen in the Example 2.
Output information is also written to the json format. This means that the status, solution status, the values of variables and shadow prices / reduced costs are exported too. This means that it is possible to export a model that has been solved and then read it again only to see the values of the variables.
For json, we use the base json package. But if ujson is available, we use that so the import / export can be really fast.
Example 1: json
A very simple example taken from the internal tests. Imagine the following problem:
from pulp import *
prob = LpProblem("test_export_dict_MIP", LpMinimize)
x = LpVariable("x", 0, 4)
y = LpVariable("y", -1, 1)
z = LpVariable("z", 0, None, LpInteger)
prob += x + 4 * y + 9 * z, "obj"
prob += x + y <= 5, "c1"
prob += x + z >= 10, "c2"
prob += -y + z == 7.5, "c3"
We can now export the problem into a dictionary:
data = prob.to_dict()
We now have a dictionary with a lot of data:
{'constraints': [{'coefficients': [{'name': 'x', 'value': 1},
{'name': 'y', 'value': 1}],
'constant': -5,
'name': 'c1',
'pi': None,
'sense': -1},
{'coefficients': [{'name': 'x', 'value': 1},
{'name': 'z', 'value': 1}],
'constant': -10,
'name': 'c2',
'pi': None,
'sense': 1},
{'coefficients': [{'name': 'y', 'value': -1},
{'name': 'z', 'value': 1}],
'constant': -7.5,
'name': 'c3',
'pi': None,
'sense': 0}],
'objective': {'coefficients': [{'name': 'x', 'value': 1},
{'name': 'y', 'value': 4},
{'name': 'z', 'value': 9}],
'name': 'obj'},
'parameters': {'name': 'test_export_dict_MIP',
'sense': 1,
'sol_status': 0,
'status': 0},
'sos1': {},
'sos2': {},
'variables': [{'cat': 'Continuous',
'dj': None,
'lowBound': 0,
'name': 'x',
'upBound': 4,
'varValue': None},
{'cat': 'Continuous',
'dj': None,
'lowBound': -1,
'name': 'y',
'upBound': 1,
'varValue': None},
{'cat': 'Integer',
'dj': None,
'lowBound': 0,
'name': 'z',
'upBound': None,
'varValue': None}]}
We can now import this dictionary:
var1, prob1 = LpProblem.from_dict(data)
var1
# {'x': x, 'y': y, 'z': z}
prob1
# test_export_dict_MIP:
# MINIMIZE
# 1*x + 4*y + 9*z + 0
# SUBJECT TO
# c1: x + y <= 5
# c2: x + z >= 10
# c3: - y + z = 7.5
# VARIABLES
# x <= 4 Continuous
# -1 <= y <= 1 Continuous
# 0 <= z Integer
As you can see we get a tuple with size 2 with: (1) a variables dictionary and (2) a PuLP model object. We can now solve that problem:
prob1.solve()
And the result will be available in our new variables:
var1['x'].value()
# 3.0
Example 1: mps
The same model:
from pulp import *
prob = LpProblem("test_export_dict_MIP", LpMinimize)
x = LpVariable("x", 0, 4)
y = LpVariable("y", -1, 1)
z = LpVariable("z", 0, None, LpInteger)
prob += x + 4 * y + 9 * z, "obj"
prob += x + y <= 5, "c1"
prob += x + z >= 10, "c2"
prob += -y + z == 7.5, "c3"
We can now export the problem into an mps file:
prob.writeMPS("test.mps")
We can now import this file:
var1, prob1 = LpProblem.fromMPS("test.mps")
var1
# {'x': x, 'y': y, 'z': z}
prob1
# test_export_dict_MIP:
# MINIMIZE
# 1*x + 4*y + 9*z + 0
# SUBJECT TO
# c1: x + y <= 5
# c2: x + z >= 10
# c3: - y + z = 7.5
# VARIABLES
# x <= 4 Continuous
# -1 <= y <= 1 Continuous
# 0 <= z Integer
The resulting tuple is exactly the same format as the previous one.
Example 2: json
We will use as example the model in A Set Partitioning Problem:
import pulp
max_tables = 5
max_table_size = 4
guests = 'A B C D E F G I J K L M N O P Q R'.split()
def happiness(table):
"""
Find the happiness of the table
- by calculating the maximum distance between the letters
"""
return abs(ord(table[0]) - ord(table[-1]))
#create list of all possible tables
possible_tables = [tuple(c) for c in pulp.allcombinations(guests,
max_table_size)]
#create a binary variable to state that a table setting is used
x = pulp.LpVariable.dicts('table', possible_tables,
lowBound = 0,
upBound = 1,
cat = pulp.LpInteger)
seating_model = pulp.LpProblem("Wedding_Seating_Model", pulp.LpMinimize)
seating_model += pulp.lpSum([happiness(table) * x[table] for table in possible_tables])
#specify the maximum number of tables
seating_model += pulp.lpSum([x[table] for table in possible_tables]) <= max_tables, \
"Maximum_number_of_tables"
#A guest must seated at one and only one table
for guest in guests:
seating_model += pulp.lpSum([x[table] for table in possible_tables
if guest in table]) == 1, "Must_seat_%s"%guest
We could directly solve the model doing:
seating_model.solve()
Instead, we are going to export it to a json file:
seating_model.to_json("seating_model.json")
And re-import it:
wedding_vars, wedding_model = LpProblem.from_json("seating_model.json")
We inspect the variables:
wedding_vars
{"table_('A',)": table_('A',), "table_('A',_'B')": table_('A',_'B'), "table_('A',_'B',_'C')": table_('A',_'B',_'C'), "table_('A',_'B',_'C',_'D')": table_('A',_'B',_'C',_'D'), "table_('A',_'B',_'C',_'E')": table_('A',_'B',_'C',_'E'), ...}
As can be seen, it is no longer a dictionary indexed by the original tuples. Unfortunately, it has become a flat dictionary with concatenated names.
We can still solve the model, though:
wedding_model.solve()
And inspect some of the values:
wedding_vars["table_('M',_'N')"].value()
# 1.0
Grouping variables
As the “Considerations” section mentions, the grouping of variables is not restored automatically. Nevertheless, by using some strict naming convention on variable names and clever parsing, one can reconstruct the original structure of the variables.
Caveats with json and pandas / numpy data types
The json module in python has some issues transforming numpy data types (e.g., np.integer). The easier way to solve this problem is to provide a custom encoding class as shown here:
import numpy as np
#(...)
class NpEncoder(json.JSONEncoder):
def default(self, obj):
if isinstance(obj, np.integer):
return int(obj)
elif isinstance(obj, np.floating):
return float(obj)
elif isinstance(obj, np.ndarray):
return obj.tolist()
else:
return super(NpEncoder, self).default(obj)
wedding_model.to_json("seating_model.json", cls=NpEncoder)
Note that this custom encoding class may not work with the ujson package. An alternative is to cast all values using int() or float() before using them in pulp.