使用来自 Scikit Optimize 的 @use_named_args

Question

I'm having a problem on using @use_named_args from Scikit Optimize . The problem is that my objective function accepts the arguments NamedTuple and I can't change this because this is the requirement in the project I'm working on. Now, I need to implement skopt for hyperparameters search and I need to use @use_named_args to decorate my objective function. How can I do it since my objective function accepting NamedTuple instead of single arguments (like the one on skopt example)? I also need to pass a fixed hyperparameters set in addition to the variable hyperparameters that I need to tune.

Below is the code I want to achieve, but I can't because I can't decorate my_objective_function with @use_named_args

from skopt.space import Real
from skopt import forest_minimize
from skopt.utils import use_named_args
from functools import partial

dim1 = Real(name='foo', low=0.0, high=1.0)
dim2 = Real(name='bar', low=0.0, high=1.0)
dim3 = Real(name='baz', low=0.0, high=1.0)

dimensions = [dim1, dim2, dim3]

class variable_params(NamedTuple):
    bar: int
    foo: int
    baz: int

class fixed_params(NamedTuple):
    bar1: int
    foo1: int
    baz1: int

# Instantiate object
variable_args = variable_params(foo=5, bar=10, baz=2)
fixed_args = fixed_params(foo1=2, bar1=3, baz1=4)


@use_named_args(dimensions=dimensions)
def my_objective_function(v_args, f_args):
    return v_args.foo ** 2 + v_args.bar ** 4 + v_args.baz ** 8 + f_args.foo1 * 2 + f_args.bar1 * 4 + f_args.baz1 * 8

#Do partial function for passing the fixed params
my_objective_function = partial(my_objective_function,f_args=fixed_args)

result = forest_minimize(
    func=my_objective_function, 
    dimensions=dimensions, 
    n_calls=20, 
    base_estimator="ET",
    random_state=4
)

Thank you!

Answer 1

You can just create a new objective function to be passed to the optimizer. It will receive the variable parameters, convert those to a named tuple and then call the original objective.

Slightly adjusting your example you get something like:

from skopt.space import Real
from skopt import forest_minimize
from skopt.utils import use_named_args
from collections import namedtuple

dim1 = Real(name='foo', low=0.0, high=1.0)
dim2 = Real(name='bar', low=0.0, high=1.0)
dim3 = Real(name='baz', low=0.0, high=1.0)

dimensions = [dim1, dim2, dim3]

VariableParams = namedtuple('VariableParams', 'foo bar baz')
FixedParams = namedtuple('FixedParams', 'foo1 bar1 baz1')

# define fixed params
fixed_args = FixedParams(foo1=2, bar1=3, baz1=4)


# objective you are not allowed to change
def my_objective_function(v_args, f_args):
    return v_args.foo ** 2 + v_args.bar ** 4 + v_args.baz ** 8 + f_args.foo1 * 2 + f_args.bar1 * 4 + f_args.baz1 * 8


# new objective passed to the optimizer
@use_named_args(dimensions)
def objective(foo, bar, baz):
    variable_args = VariableParams(foo, bar, baz)
    return my_objective_function(variable_args, fixed_args)


# run search with new objective
result = forest_minimize(
    func=objective,
    dimensions=dimensions,
    n_calls=10
)