def test_soft_trimmed_degenerated(self):
"""Tests several possible degenerated cases."""
with self.assertRaises(ValueError):
losses.SoftTrimmedRegressionLoss(start_quantile=0.6,
end_quantile=0.1)
loss_fn = losses.SoftTrimmedRegressionLoss(start_quantile=0.1,
end_quantile=0.3)
weights, index = loss_fn._get_target_weights_and_indices()
self.assertLen(weights, 3)
self.assertEqual(index, 1)
loss_fn = losses.SoftTrimmedRegressionLoss(start_quantile=0.0,
end_quantile=0.2)
weights, index = loss_fn._get_target_weights_and_indices()
self.assertLen(weights, 2)
self.assertEqual(index, 0)
loss_fn = losses.SoftTrimmedRegressionLoss(start_quantile=0.4,
end_quantile=1.0)
weights, index = loss_fn._get_target_weights_and_indices()
self.assertLen(weights, 2)
self.assertEqual(index, 1)
loss_fn = losses.SoftTrimmedRegressionLoss(start_quantile=0.0,
end_quantile=1.0)
weights, index = loss_fn._get_target_weights_and_indices()
self.assertLen(weights, 1)
self.assertEqual(index, 0)
def test_soft_trimmed_degenerated(self, start, end):
"""Tests several possible degenerated cases."""
loss_fn = losses.SoftTrimmedRegressionLoss(start_quantile=start,
end_quantile=end)
try:
loss_fn(self._y_true, self._y_pred)
except ValueError:
self.fail(
'SoftTrimmedRegressionLoss raised ValueError unexpectedly!')
def test_soft_trimmed(self, start, end, power):
loss_fn = losses.SoftTrimmedRegressionLoss(start_quantile=start,
end_quantile=end,
power=power)
loss = loss_fn(self._y_true, self._y_pred)
start_index = int(start * self._num_points)
end_index = int(end * self._num_points)
selected = tf.pow(self._values[start_index:end_index], power)
expected_loss = tf.math.reduce_mean(selected)
self.assertAllClose(loss, expected_loss, 0.2, 0.2)