def testBijectorForward(self):
bijector = categorical_to_discrete.CategoricalToDiscrete(
map_values=[0.1, 0.2, 0.3, 0.4], validate_args=True)
self.assertAllClose([[0.1, 0.2, 0.3, 0.4], [0.4, 0.3, 0.2, 0.1]],
self.evaluate(
bijector.forward([[0, 1, 2, 3], [3, 2, 1,
0]])))
def testInverseLogDetJacobian(self):
bijector = categorical_to_discrete.CategoricalToDiscrete(
map_values=[0.1, 0.2], validate_args=True)
self.assertAllClose(
0,
self.evaluate(
bijector.inverse_log_det_jacobian([0.1, 0.2], event_ndims=0)))
def testBijectorInverse(self):
bijector = categorical_to_discrete.CategoricalToDiscrete(
map_values=[0.1, 0.2, 0.3, 0.4], validate_args=True)
self.assertAllEqual([[3, 3, 3], [0, 1, 2]],
self.evaluate(
bijector.inverse([[0.400001, 0.4, 0.399999],
[0.1, 0.2, 0.3]])))
def testVariableGradients(self):
map_values = tf.Variable([0.3, 0.5])
b = categorical_to_discrete.CategoricalToDiscrete(
map_values=map_values, validate_args=True)
with tf.GradientTape() as tape:
y = tf.reduce_sum(b.forward([0, 1]))
grads = tape.gradient(y, [map_values])
self.assertAllNotNone(grads)
def testModifiedMapValuesIncreasingAssertion(self):
map_values = tf.Variable([0.1, 0.2])
b = categorical_to_discrete.CategoricalToDiscrete(map_values=map_values,
validate_args=True)
self.evaluate(map_values.initializer)
with self.assertRaisesOpError('map_values is not strictly increasing.'):
with tf.control_dependencies([map_values.assign([0.2, 0.1])]):
self.evaluate(b.forward([0, 1]))
def testBijectiveAndFinite16bit(self):
x = np.arange(100).astype(np.int32)
y = np.logspace(-5, 4, 100).astype(np.float16)
bijector = categorical_to_discrete.CategoricalToDiscrete(map_values=y)
bijector_test_util.assert_bijective_and_finite(bijector,
x,
y,
eval_func=self.evaluate,
event_ndims=0)
def testBijectorInverseValueNotFoundRaises(self):
with self.assertRaisesOpError('inverse value not found'):
bijector = categorical_to_discrete.CategoricalToDiscrete(
map_values=[0.1, 0.2, 0.3, 0.4], validate_args=True)
self.evaluate(bijector.inverse([0.21, 0.4]))
def testBijectorForwardOutOfBoundIndicesRaises(self):
with self.assertRaisesOpError('indices out of bound'):
bijector = categorical_to_discrete.CategoricalToDiscrete(
map_values=[0.1, 0.2, 0.3, 0.4], validate_args=True)
self.evaluate(bijector.forward([5]))
def testMapValuesSizeZeroRaises(self):
with self.assertRaisesWithPredicateMatch(
ValueError, 'Size of map_values must be greater than 0'):
bijector = categorical_to_discrete.CategoricalToDiscrete(
map_values=[], validate_args=True)
self.evaluate(bijector.map_values)
def testMapValuesRankNotEqualToOneRaises(self):
with self.assertRaisesWithPredicateMatch(
ValueError, 'Rank of map_values must be 1'):
bijector = categorical_to_discrete.CategoricalToDiscrete(
map_values=[[1, 2], [3, 4]], validate_args=True)
self.evaluate(bijector.map_values)
def testUnsortedValuesRaises(self):
with self.assertRaisesOpError('map_values is not strictly increasing'):
bijector = categorical_to_discrete.CategoricalToDiscrete(
map_values=[1, 3, 2], validate_args=True)
self.evaluate(bijector.forward([0, 1, 2]))
def _func(map_values):
b = categorical_to_discrete.CategoricalToDiscrete(
map_values=map_values, validate_args=True)
return tf.reduce_sum(b.forward([0, 1]))
