def test_softmax_loss(self):
scores = [[1., 3., 2.], [1., 2., 3.], [1., 2., 3.]]
labels = [[0., 0., 1.], [0., 0., 2.], [0., 0., 0.]]
weights = [[2.], [1.], [1.]]
loss = losses.SoftmaxLoss()
self.assertAlmostEqual(
loss(labels, scores).numpy(),
-(ln(_softmax(scores[0])[2]) + ln(_softmax(scores[1])[2]) * 2.) /
3.,
places=5)
self.assertAlmostEqual(loss(labels, scores, weights).numpy(),
-(ln(_softmax(scores[0])[2]) * 2. +
ln(_softmax(scores[1])[2]) * 2. * 1.) / 3.,
places=5)
# Test LambdaWeight.
rank_discount_fn = lambda r: 1. / tf.math.log1p(r)
lambda_weight = losses.DCGLambdaWeight(
rank_discount_fn=rank_discount_fn)
loss = losses.SoftmaxLoss(lambda_weight=lambda_weight)
self.assertAlmostEqual(
loss(labels, scores).numpy(),
-(ln(_softmax(scores[0])[2]) / ln(1. + 2.) +
ln(_softmax(scores[1])[2]) * 2. / ln(1. + 1.)) / 3.,
places=5)
def setUp(self):
super(FunctionalRankingModelTest, self).setUp()
self.context_feature_columns = _context_feature_columns()
self.example_feature_columns = _example_feature_columns()
self.optimizer = tf.keras.optimizers.Adagrad()
self.loss = losses.SoftmaxLoss()
self.metrics = [metrics.NDCGMetric("ndcg_5", topn=5)]
def test_softmax_loss_with_invalid_labels(self):
scores = [[1., 3., 2.]]
labels = [[0., -1., 1.]]
loss = losses.SoftmaxLoss()
self.assertAlmostEqual(
loss(labels, scores).numpy(), -(ln(_softmax([1, 2])[1])), places=5)
def test_model_compile_keras(self):
# Specify the training configuration (optimizer, loss, metrics).
optimizer = tf.keras.optimizers.RMSprop()
loss = losses.SoftmaxLoss()
eval_metrics = [metrics.NDCGMetric("ndcg_5", topn=5)]
ranker = model.create_keras_model(network=self.network,
loss=loss,
metrics=eval_metrics,
optimizer=optimizer,
size_feature_name=None)
self.assertIs(ranker.optimizer, optimizer)
self.assertIs(ranker.loss, loss)
def test_listwise_losses_are_serializable(self):
self.assertIsLossSerializable(
losses.SoftmaxLoss(lambda_weight=self._lambda_weight))
self.assertIsLossSerializable(
losses.ListMLELoss(lambda_weight=self._lambda_weight))
self.assertIsLossSerializable(
losses.ApproxMRRLoss(lambda_weight=self._lambda_weight))
self.assertIsLossSerializable(
losses.ApproxNDCGLoss(lambda_weight=self._lambda_weight))
# TODO: Debug assertIsLossSerializable for Gumbel loss. Right now,
# the loss values got from obj and the deserialized don't match exactly.
self.assertIsSerializable(losses.GumbelApproxNDCGLoss(seed=1))
