def test_softmax_loss(self):
with tf.Graph().as_default():
scores = [[1., 3., 2.], [1., 2., 3.], [1., 2., 3.]]
labels = [[0., 0., 1.], [0., 0., 2.], [0., 0., 0.]]
weights = [[2.], [1.], [1.]]
with self.cached_session():
self.assertAlmostEqual(
ranking_losses._softmax_loss(labels, scores).eval(),
-(math.log(_softmax(scores[0])[2]) +
math.log(_softmax(scores[1])[2]) * 2.) / 2.,
places=5)
self.assertAlmostEqual(
ranking_losses._softmax_loss(labels, scores,
weights).eval(),
-(math.log(_softmax(scores[0])[2]) * 2. +
math.log(_softmax(scores[1])[2]) * 2. * 1.) / 2.,
places=5)
# Test LambdaWeight.
lambda_weight = ranking_losses.DCGLambdaWeight(
rank_discount_fn=lambda r: 1. / tf.math.log1p(r))
self.assertAlmostEqual(
ranking_losses._softmax_loss(
labels, scores, lambda_weight=lambda_weight).eval(),
-(math.log(_softmax(scores[0])[2]) / math.log(1. + 2.) +
math.log(_softmax(scores[1])[2]) * 2. /
math.log(1. + 1.)) / 2.,
places=5)
def test_softmax_loss_with_invalid_labels(self):
scores = [[1., 3., 2.]]
labels = [[0., -1., 1.]]
with self.cached_session():
self.assertAlmostEqual(
ranking_losses._softmax_loss(labels, scores).eval(),
-(math.log(_softmax([1, 2])[1])),
places=5)