def test_minDCF():
from speechbrain.utils.metric_stats import minDCF
positive_scores = torch.tensor([0.1, 0.2, 0.3])
negative_scores = torch.tensor([0.4, 0.5, 0.6])
min_dcf, threshold = minDCF(positive_scores, negative_scores)
assert (0.01 - min_dcf) < 1e-4
assert threshold >= 0.6
positive_scores = torch.tensor([0.4, 0.5, 0.6])
negative_scores = torch.tensor([0.1, 0.2, 0.3])
min_dcf, threshold = minDCF(positive_scores, negative_scores)
assert min_dcf == 0
assert threshold > 0.3 and threshold < 0.4
def verification_performance(scores_plda):
"""Computes the Equal Error Rate give the PLDA scores"""
# Create ids, labels, and scoring list for EER evaluation
ids = []
labels = []
positive_scores = []
negative_scores = []
for line in open(veri_file_path):
lab = int(line.split(" ")[0].rstrip().split(".")[0].strip())
enrol_id = line.split(" ")[1].rstrip().split(".")[0].strip()
test_id = line.split(" ")[2].rstrip().split(".")[0].strip()
# Assuming enrol_id and test_id are unique
i = int(numpy.where(scores_plda.modelset == enrol_id)[0][0])
j = int(numpy.where(scores_plda.segset == test_id)[0][0])
s = float(scores_plda.scoremat[i, j])
labels.append(lab)
ids.append(enrol_id + "<>" + test_id)
if lab == 1:
positive_scores.append(s)
else:
negative_scores.append(s)
# Clean variable
del scores_plda
# Final EER computation
eer, th = EER(torch.tensor(positive_scores), torch.tensor(negative_scores))
min_dcf, th = minDCF(torch.tensor(positive_scores),
torch.tensor(negative_scores))
return eer, min_dcf
def verification_performance(self,):
""" Computes the EER using the standard voxceleb test split
"""
# Computing enrollment and test embeddings
print("Computing enroll/test embeddings...")
self.enrol_dict = self.compute_embeddings_loop(enrol_dataloader)
self.test_dict = self.compute_embeddings_loop(test_dataloader)
print("Computing EER..")
# Reading standard verification split
gt_file = os.path.join(
self.hparams.data_folder, "meta", "veri_test.txt"
)
with open(gt_file) as f:
veri_test = [line.rstrip() for line in f]
positive_scores, negative_scores = self.get_verification_scores(
veri_test
)
del self.enrol_dict, self.test_dict
eer, th = EER(
torch.tensor(positive_scores), torch.tensor(negative_scores)
)
min_dcf, th = minDCF(
torch.tensor(positive_scores), torch.tensor(negative_scores)
)
return eer * 100, min_dcf
# Computing enrollment and test embeddings
logger.info("Computing enroll/test embeddings...")
# First run
enrol_dict = compute_embedding_loop(enrol_dataloader)
test_dict = compute_embedding_loop(test_dataloader)
# Second run (normalization stats are more stable)
enrol_dict = compute_embedding_loop(enrol_dataloader)
test_dict = compute_embedding_loop(test_dataloader)
if "score_norm" in params:
train_dict = compute_embedding_loop(train_dataloader)
# Compute the EER
logger.info("Computing EER..")
# Reading standard verification split
with open(veri_file_path) as f:
veri_test = [line.rstrip() for line in f]
positive_scores, negative_scores = get_verification_scores(veri_test)
del enrol_dict, test_dict
eer, th = EER(torch.tensor(positive_scores), torch.tensor(negative_scores))
logger.info("EER(%%)=%f", eer * 100)
min_dcf, th = minDCF(torch.tensor(positive_scores),
torch.tensor(negative_scores))
logger.info("minDCF=%f", min_dcf * 100)