angleFinetuning/example_codes.py at master · lrast/angleFinetuning · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
# Example codes generated by KL minimization. Used to develop active learning methods
import numpy as np
import cvxpy as cp

from scipy.optimize import minimize, LinearConstraint
from scipy.stats import entropy


def min_kl_cvxpy(ps, Cs, C0, direction='forward', return_lagrange=False):
    """ solve for KL optimal probability vector """
    n = ps.shape[0]

    qs = cp.Variable(n)

    if direction == 'forward':
        objective_fn = cp.sum(cp.rel_entr(ps, qs))
    if direction == 'reverse':
        objective_fn = cp.sum(cp.rel_entr(qs, ps))

    constraints = [
                    cp.sum(qs) == 1,
                    qs @ Cs == C0
                  ]

    prob = cp.Problem(cp.Minimize(objective_fn), constraints)
    optimal = prob.solve()
    if optimal == float('inf') or optimal == -float('inf'):
        raise Exception("Unsolved optimization")

    if return_lagrange:
        return qs.value, (constraints[0].dual_value, constraints[1].dual_value)

    return qs.value


def min_kl_scipy(ps, Cs, C0, direction='forward'):
    """ solve for KL optimal probability vector
        scipy solvers don't seem to perform as well as cvxpy
    """
    n = ps.shape[0]

    if direction == 'forward':
        to_minimize = lambda qs: entropy(ps, qs)
    if direction == 'reverse':
        to_minimize = lambda qs: entropy(qs, ps)

    constraints = LinearConstraint([Cs, np.ones(n)], [C0, 1], [C0, 1])

    initial_guess = np.ones(n) / n
    res = minimize(to_minimize, initial_guess, method='trust-constr',
                   constraints=constraints)
    return res