def test_np_inverse_softmax(self):
batch_size, nclasses = [4, 3]
logits_orig = np.random.rand(batch_size, nclasses)
probs_orig = scipy.special.softmax(logits_orig, axis=-1)
logits_new = uq_utils.np_inverse_softmax(probs_orig)
probs_new = scipy.special.softmax(logits_new, axis=-1)
self.assertAllClose(probs_orig, probs_new)
def apply_temperature_scaling(temperature, probs):
"""Apply temperature scaling to an array of probabilities.
Args:
temperature: Floating point temperature.
probs: Array of probabilities with probabilities over axis=-1.
Returns:
Temperature-scaled probabilities; same shape as input probs.
"""
logits_t = uq_utils.np_inverse_softmax(probs).T / temperature
return scipy.special.softmax(logits_t.T, axis=-1)
def run(prediction_path):
"""Run temperature scaling."""
stats = array_utils.load_stats_from_tfrecords(prediction_path)
probs = stats['probs'].astype(np.float32)
labels = stats['labels'].astype(np.int32)
if len(labels.shape) > 1:
labels = np.squeeze(labels, -1)
if probs.shape[0] > NUM_EXAMPLES:
probs = probs[:NUM_EXAMPLES, :]
labels = labels[:NUM_EXAMPLES]
probs = metrics_lib.soften_probabilities(probs=probs)
logits = uq_utils.np_inverse_softmax(probs)
temp = calibration_lib.find_scaling_temperature(labels, logits)
with gfile.GFile(
os.path.join(os.path.dirname(prediction_path),
'temperature_hparam.json'), 'w') as fh:
fh.write(json.dumps({'temperature': temp}))
