def eer_bob_rocch(impostors, genuines, is_sorted=False):
"""
EER function that wraps LDIAP's bob.measure toolkit EER function.
:param impostors: An iterable of impostor scores
:param genuines: An iterable of genuine scores
:param is_sorted: Bool indicating if iterable is sorted
:return: EER
"""
return eer_rocch(impostors, genuines)
def bootstrap_sorted_bob(impostors, genuines, m=1):
eers = [0] * m
impostors = np.array(sorted(impostors))
genuines = np.array(sorted(genuines))
for c_iter in range(m):
imps = bootstrap_draw_sorted(impostors)
gens = bootstrap_draw_sorted(genuines)
eers[c_iter] = eer_rocch(imps, gens)
return eers
def bootstrap_naive_bob(impostors, genuines, m=1):
eers = [0] * m
max_rounds = 20
i = 0
# It is more time efficient to generate more sample draws in one call to the numpy api. However,
# this comes at considerable cost in memory (10,000 * 100,000 32-bit scores is 4 GB for example).
# Therefore, we try to strike a balance between memory and computation.
for i in range(int(m / max_rounds)):
imps = np.random.choice(impostors, (max_rounds, len(impostors)))
gens = np.random.choice(genuines, (max_rounds, len(genuines)))
for c_iter in range(max_rounds):
eers[c_iter + i * max_rounds] = eer_rocch(imps[c_iter, :],
gens[c_iter, :])
imps = np.random.choice(impostors, (m % max_rounds, len(impostors)))
gens = np.random.choice(genuines, (m % max_rounds, len(genuines)))
for c_iter in range(m % max_rounds):
eers[c_iter + (i + 1) * max_rounds] = eer_rocch(
imps[c_iter, :], gens[c_iter, :])
return eers
def bootstrap_naive_rocch(impostors, genuines):
imps = bootstrap_draw_sorted(impostors)
gens = bootstrap_draw_sorted(genuines)
return eer_rocch(imps, gens)
def bootstrap_naive_bob_once(imps, gens):
return eer_rocch(imps, gens)