def predict(self, logits, temperature=None):
"""Scales logits based on the temperature, returns calibrated probabilities.
Args:
logits: raw (non-normalized) predictions that a classification model
generates, shape=(num_examples, num_classes)
temperature: temperature to scale logits by
Returns:
calibrated softmax probabilities shape=(num_samples, num_classes)
"""
if not temperature:
return utils.to_softmax(logits / self.temperature)
else:
return utils.to_softmax(logits / temperature)
def fit(self, logits, one_hot_labels):
"""Fit the isotonic regression model.
Args:
logits: raw (non-normalized) predictions that a classification model
generates, shape=(num_examples, num_classes)
one_hot_labels: one-hot-encoding of true labels, shape=(num_examples,
num_classes)
"""
assert logits.shape[1] == self.num_classes
assert logits.shape == one_hot_labels.shape
softmax_probabilities = utils.to_softmax(logits)
for i in range(self.num_classes):
self.ir_per_class[i].fit(softmax_probabilities[:, i],
one_hot_labels[:, i])
def predict(self, logits):
"""Predict new softmax probabilities from logit scores.
Uses linear interpolation in underlying scikit learn call.
Args:
logits: raw (non-normalized) predictions that a classification model
generates, shape=(num_examples, num_classes)
Returns:
calibrated softmax probabilities, shape = (num_examples, num_classes)
"""
assert logits.shape[1] == self.num_classes
input_probabilities = utils.to_softmax(logits)
new_probabilities = np.ones(np.shape(input_probabilities))
for i in range(self.num_classes):
new_probabilities[:, i] = self.ir_per_class[i].predict(
input_probabilities[:, i])
# normalize each row of the probability vector if in multiclass setting
if self.num_classes > 1:
row_sums = np.sum(new_probabilities, axis=1)
return new_probabilities / row_sums[:, np.newaxis]
else:
return new_probabilities