def test_make_loss_metric_fn(self):
scores = [[1., 3., 2.], [1., 2., 3.]]
labels = [[0., 0., 1.], [0., 0., 2.]]
weights = [[2.], [1.]]
weights_feature_name = 'weights'
features = {
weights_feature_name: weights,
}
m = ranking_losses.make_loss_metric_fn(
ranking_losses.RankingLossKey.SOFTMAX_LOSS)
m_w = ranking_losses.make_loss_metric_fn(
ranking_losses.RankingLossKey.SOFTMAX_LOSS,
weights_feature_name=weights_feature_name)
self._check_metrics([
(m(labels, scores,
features), -(math.log(_softmax(scores[0])[2]) +
math.log(_softmax(scores[1])[2]) * 2.) / 3.),
(m_w(labels, scores,
features), -(math.log(_softmax(scores[0])[2]) * 2. +
math.log(_softmax(scores[1])[2]) * 2. * 1.) / 4.),
])
# Value of loss metric is the same as loss with MEAN reduction.
with self.cached_session():
loss_fn_mean = ranking_losses.make_loss_fn(
ranking_losses.RankingLossKey.SOFTMAX_LOSS,
reduction=tf.compat.v1.losses.Reduction.MEAN)
loss_mean = loss_fn_mean(labels, scores, features).eval()
self._check_metrics([
(m(labels, scores, features), loss_mean),
])
def _get_loss_metric_pair(key, weight=None):
"""Helper function to construct metric name and function for a loss."""
name = "".join([
"metric/",
"weighted_" if weight else "",
key,
])
return name, losses.make_loss_metric_fn(key, weights_feature_name=weight)