def negloglikelihood(y, p_tab):
"""Compute the negative log-likelihood.
Args:
y: one-hot encoding of the true classes, size (?, num_classes)
p_tab: numpy array, size (?, num_classes, num_samples)
containing the output predicted probabilities
Returns:
neglog: negative log likelihood, along the iterations
numpy vector of size num_samples
"""
p_mean = util.cummean(p_tab[y.astype(np.bool), :], axis=1)
neglog = -np.mean(np.log(p_mean), axis=0)
return neglog
def accuracy(y, p_tab):
"""Compute the accuracy.
Args:
y: one-hot encoding of the true classes, size (?, num_classes)
p_tab: numpy array, size (?, num_classes, num_samples)
containing the output predicted probabilities
Returns:
acc: accuracy along the iterations, numpy vector of size num_samples
"""
class_pred = np.argmax(util.cummean(p_tab, axis=2), axis=1)
argmax_y = np.argmax(y, axis=1)
acc = np.apply_along_axis(lambda x: np.mean(x == argmax_y),
axis=0,
arr=class_pred)
return acc
def brier_score(y, p_tab):
"""Compute the Brier score.
Brier Score: see
https://www.stat.washington.edu/raftery/Research/PDF/Gneiting2007jasa.pdf,
page 363, Example 1
Args:
y: one-hot encoding of the true classes, size (?, num_classes)
p_tab: numpy array, size (?, num_classes, num_samples)
containing the output predicted probabilities
Returns:
bs: Brier score along the iteration, vector of size num_samples.
"""
p_cummean = util.cummean(p_tab, axis=2)
y_repeated = np.repeat(y[:, :, np.newaxis], p_tab.shape[2], axis=2)
bs = np.mean(np.power(p_cummean - y_repeated, 2), axis=(0, 1))
return bs
