def negative_log_likelihood(y_true, y_pred):
"""
Negative log-likelihood or negative log-probability loss/metric.
Reference: Evaluating Predictive Uncertainty Challenge, Quiñonero-Candela et al, 2006.
It sums over classes: log(y_pred) for true class and log(1.0 - pred) for not true class, and then takes average across samples.
"""
y_pred = K.clip(y_pred, K.epsilon(), 1.0 - K.epsilon())
return -K.mean(K.sum(y_true * K.log(y_pred) +
(1.0 - y_true) * K.log(1.0 - y_pred),
axis=-1),
axis=-1)
def nll(y_true, y_pred):
#if variance_logits:
# variance_tensor = K.exp(variance_tensor)
return 0.5 * K.mean(
K.log(variance_tensor + epsilon) + K.square(y_true - y_pred) /
(variance_tensor + epsilon))
def entropy(y_true, y_pred):
"""
Standard entropy over class probabilities.
It sums y_pred * K.log(y_pred + epsilon) over class probabilities, and then takes average over samples
"""
return K.mean(-K.sum(y_pred * K.log(y_pred + K.epsilon()), axis=-1),
axis=-1)
def beta_nll(y_true, y_pred):
#if variance_logits:
# variance_tensor = K.exp(variance_tensor)
beta_sigma_sq = K.stop_gradient(K.pow(variance_tensor, beta))
return 0.5 * K.mean(
beta_sigma_sq *
(K.log(variance_tensor + epsilon) + K.square(y_true - y_pred) /
(variance_tensor + epsilon)))
def nll(y_true, y_pred):
return K.mean(
K.log(2.0 * spread_tensor + epsilon) + K.abs(y_true - y_pred) /
(spread_tensor + epsilon))
def log_prior_prob(self, w):
return K.log(self.prior_pi_1 *
gaussian.probability(w, 0.0, self.prior_sigma_1) +
self.prior_pi_2 *
gaussian.probability(w, 0.0, self.prior_sigma_2))
def log_probability(x, mu, sigma):
return NegHalfLog2PI - K.log(sigma) - 0.5 * K.square((x - mu) / sigma)
def log_probability(self, x):
return NegHalfLog2PI - K.log(self.std) - 0.5 * K.square(
(x - self.mean) / self.std)