def _testRegularizationPenalty(var, l1_scale, l2_scale,
l1_penalty_truth, l2_penalty_truth):
"""Regularization test for a given variable value"""
with tf.Graph().as_default() as g:
tf.Variable(var,
trainable=True) # Register variable in the graph
# Get L1 loss
l1_hparams = tf.contrib.training.HParams(l1=l1_scale, l2=None)
l1_penalty = compute_regularization_penalty(l1_hparams)
# Get L2 loss
l2_hparams = tf.contrib.training.HParams(l1=None, l2=l2_scale)
l2_penalty = compute_regularization_penalty(l2_hparams)
# Get elastic net loss
elastic_net_hparams = tf.contrib.training.HParams(l1=l1_scale,
l2=l2_scale)
elastic_net_penalty = compute_regularization_penalty(
elastic_net_hparams)
with tf.Session(graph=g) as sess:
sess.run([tf.global_variables_initializer()])
self.assertAlmostEqual(l1_penalty.eval(), l1_penalty_truth)
self.assertAlmostEqual(l2_penalty.eval(), l2_penalty_truth)
self.assertAlmostEqual(elastic_net_penalty.eval(),
l1_penalty_truth + l2_penalty_truth)
def compute_loss(hparams, scores, labels, group_size, weight):
""" Computes ranking/classification loss with regularization """
if weight is None:
raise ValueError("weight should not be None")
return compute_rank_clf_loss(
hparams, scores, labels, group_size, weight
) + tf.reduce_mean(weight) * compute_regularization_penalty(hparams)
def _testRegularizationPenalty(self, var, l1_scale, l2_scale,
l1_penalty_truth, l2_penalty_truth):
"""Regularization test for a given variable value"""
# Get L1 loss
l1_penalty = compute_regularization_penalty(l1_scale, None, [var])
# Get L2 loss
l2_penalty = compute_regularization_penalty(None, l2_scale, [var])
# Get elastic net loss
elastic_net_penalty = compute_regularization_penalty(
l1_scale, l2_scale, [var])
self.assertAlmostEqual(l1_penalty.numpy(),
l1_penalty_truth,
places=self.places)
self.assertAlmostEqual(l2_penalty.numpy(),
l2_penalty_truth,
places=self.places)
self.assertAlmostEqual(elastic_net_penalty.numpy(),
l1_penalty_truth + l2_penalty_truth,
places=self.places)