def testProjection(self):
"""Create a partial monotone linear layer and check the projection."""
input_dim = 10
is_monotone = [True, False] * 5
input_placeholder = array_ops.placeholder(
dtype=dtypes.float32, shape=[None, input_dim])
# We set the initial_weight_mean to -10.0. After projection, we expect
# elements corresponding to monotonic input becomes 0.
packed_results = monotone_linear_layers.monotone_linear_layer(
input_placeholder,
input_dim=input_dim,
output_dim=2,
is_monotone=is_monotone,
init_weight_mean=-10.0,
init_weight_stddev=0.0)
(_, weight_tensor, projection_op, _) = packed_results
# The weight is in shape (output_dim, input_dim).
expected_pre_projection_weight = [[-10.0] * 10] * 2
expected_projected_weight = [[0.0, -10.0] * 5] * 2
with self.test_session() as sess:
sess.run(variables.global_variables_initializer())
pre_projection_weight = sess.run(weight_tensor)
sess.run(projection_op)
projected_weight = sess.run(weight_tensor)
self.assertAllClose(expected_pre_projection_weight, pre_projection_weight)
self.assertAllClose(expected_projected_weight, projected_weight)
def testNormalizationProjection(self):
"""Test projection when l1 normalization is requested."""
input_dim = 10
is_monotone = [True, False] * 5
input_placeholder = tf.compat.v1.placeholder(
dtype=tf.float32, shape=[None, input_dim])
# We set the initial_weight_mean to -10.0. After projection, we expect
# elements corresponding to monotonic input becomes 0.
packed_results = monotone_linear_layers.monotone_linear_layer(
input_placeholder,
input_dim=input_dim,
output_dim=2,
is_monotone=is_monotone,
init_weight_mean=-10.0,
init_weight_stddev=0.0,
add_bias=False,
normalization_order=1,
)
(_, weight_tensor, projection_op, _) = packed_results
# The weight is in shape (output_dim, input_dim).
expected_pre_projection_weight = [[-10.0] * 10] * 2
expected_projected_weight = [[0.0, -0.2] * 5] * 2
with self.session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
pre_projection_weight = sess.run(weight_tensor)
sess.run(projection_op)
projected_weight = sess.run(weight_tensor)
self.assertAllClose(expected_pre_projection_weight, pre_projection_weight)
self.assertAllClose(expected_projected_weight, projected_weight)
def testNoRegularizationExpectsNone(self):
"""Create a monotone linear layer and check no regularization."""
input_dim = 10
input_placeholder = array_ops.placeholder(
dtype=dtypes.float32, shape=[None, input_dim])
# We set the initial_weight_mean to -10.0.
(_, _, _, regularization) = monotone_linear_layers.monotone_linear_layer(
input_placeholder,
input_dim=input_dim,
output_dim=2,
init_weight_mean=-10.0,
init_weight_stddev=0.0)
self.assertIsNone(regularization)
def testRegularization(self):
"""Create a monotone linear layer and check regularization."""
input_dim = 10
input_placeholder = array_ops.placeholder(
dtype=dtypes.float32, shape=[None, input_dim])
# We set the initial_weight_mean to -10.0.
(_, _, _, regularization) = monotone_linear_layers.monotone_linear_layer(
input_placeholder,
input_dim=input_dim,
output_dim=2,
init_weight_mean=-10.0,
init_weight_stddev=0.0,
l1_reg=0.1,
l2_reg=0.1)
with self.test_session() as sess:
sess.run(variables.global_variables_initializer())
self.assertAlmostEqual(220.0, sess.run(regularization), delta=1e-5)
def testEvaluationWithZeroBias(self):
"""Create a partial monotone linear layer and check evaluation."""
input_placeholder = array_ops.placeholder(
dtype=dtypes.float32, shape=[None, 3])
input_tensor = [[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]
sum_input_tensor = [[2.0, 0.0, 0.0], [0.0, 2.0, 0.0], [0.0, 0.0, 2.0]]
# Check linearity of the output tensor.
# f(input_tensor + input_tensor) = 2 * f(input_tensor)
# since the bias is 0.
packed_results = monotone_linear_layers.monotone_linear_layer(
input_placeholder, input_dim=3, output_dim=5, init_bias=0.0)
(output_tensor, _, _, _) = packed_results
with self.test_session() as sess:
sess.run(variables.global_variables_initializer())
# Check linearity of the output tensor.
# f(input_tensor + input_tensor) = 2 * f(input_tensor)
# since the bias is 0.
output_val = sess.run(
output_tensor, feed_dict={input_placeholder: input_tensor})
sum_output_val = sess.run(
output_tensor, feed_dict={input_placeholder: sum_input_tensor})
expected_sum_output_val = 2 * output_val
self.assertAllClose(expected_sum_output_val, sum_output_val)
def testEvaluationWithDefaultBias(self):
"""Create a partial monotone linear layer and check the bias."""
input_dim = 10
input_placeholder = array_ops.placeholder(
dtype=dtypes.float32, shape=[None, input_dim])
# Monotone linear layers contain random weights and for this input_tensor
# we expect 0 as an output on "average". In order to control randomness, we
# set the standard deviation exactly zero.
input_tensor = [[0.5] * input_dim]
expected_output_val = [[0.0]]
packed_results = monotone_linear_layers.monotone_linear_layer(
input_placeholder,
input_dim=input_dim,
output_dim=1,
init_weight_stddev=0.0)
(output_tensor, _, _, _) = packed_results
with self.test_session() as sess:
sess.run(variables.global_variables_initializer())
# Check linearity of the output tensor.
# f(input_tensor + input_tensor) = 2 * f(input_tensor)
# since the bias is 0.
output_val = sess.run(
output_tensor, feed_dict={input_placeholder: input_tensor})
self.assertAllClose(expected_output_val, output_val)