def min_cllr(self, ptar, with_pav=False):
cllrs = []
if with_pav:
if self.pav is None:
raise Exception(
"To extract min_cllr using PAV you need to call DET with pav=True"
)
llrs, ntar, nnon = self.pav.llrs()
for p in np.atleast_1d(ptar):
logitPost = llrs + logit(p)
Ctar, Cnon = softplus(-logitPost), softplus(logitPost)
min_cllr = p * (Ctar[ntar != 0] @ ntar[ntar != 0]) / ntar.sum(
) + (1 - p) * (Cnon[nnon != 0] @ nnon[nnon != 0]) / nnon.sum()
min_cllr /= -p * np.log(p) - (1 - p) * np.log(1 - p)
cllrs.append(min_cllr)
else:
for p in np.atleast_1d(ptar):
cal = logregCal(self.tar, self.non, p)
tar_cal, non_cal = cal(self.tar), cal(self.non)
cllrs.append(cross_entropy(tar_cal, non_cal, p))
return np.array(cllrs).squeeze()
def init_with_logreg(self, scores, same_spk, valid, ptar, std_for_mats=0):
scores = scores[valid]
labels = same_spk[valid]
tar = scores[labels==1]
non = scores[labels==0]
a, b = calibration.logregCal(tar, non, ptar, return_params=True)
if self.is_durdep:
# Dur-dep calibration (might also have an si-dependent stage)
self.durdep_alpha.init_with_constant(a, std_for_mats)
self.durdep_beta.init_with_constant(b, std_for_mats)
if self.is_sidep:
self.sidep_alpha.init_with_constant(1.0, std_for_mats)
self.sidep_beta.init_with_constant(0.0, std_for_mats)
elif not self.is_sidep:
# Global calibration
utils.replace_state_dict(self, {'alpha': a, 'beta': b})
else:
# Only si-dependent cal
self.sidep_alpha.init_with_constant(a, std_for_mats)
self.sidep_beta.init_with_constant(b, std_for_mats)
def test_model(model,
data_dict,
model_name,
ptar,
loss_type,
print_min_loss=False):
# Compute the loss for each of the devsets. The validation loss will be the average over all devsets
av_loss = 0
for name, info in data_dict.items():
mask = info['mask']
scores = evaluate(model, info['dataset'], info['emap'], info['tmap'])
# Since the key mask was created to be aligned to the emap and tmap, the score_mat
# is already aligned to this mask
loss, llrs, labels = compute_loss(scores.score_mat,
mask=mask,
ptar=ptar,
loss_type=loss_type,
return_info=True)
if print_min_loss:
tar = llrs[labels == 1].detach().cpu().numpy()
non = llrs[labels == 0].detach().cpu().numpy()
cal = logregCal(tar, non, ptar)
min_loss = cross_entropy(cal(tar), cal(non), ptar)
print("%s, loss on dev set %s = %f min = %f" %
(model_name, name, loss, min_loss))
else:
print("%s, loss on dev set %s = %f" % (model_name, name, loss))
av_loss += loss
av_loss /= len(data_dict)
return av_loss