def testSampleWithIntegerOutcomes(self):
outcomes = self._build_tensor([1, 2])
probs = self._build_tensor([0.2, 0.8])
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
samples = self.evaluate(dist.sample(5000, seed=1234))
self.assertAllClose(np.mean(samples), dist.mean(), atol=0.1)
self.assertAllClose(np.std(samples), dist.stddev(), atol=0.1)
def testOutcomesNotStrictlyIncreasingRaises(self):
outcomes = self._build_tensor([1.0, 1.0, 2.0, 2.0])
probs = self._build_tensor([0.25, 0.25, 0.25, 0.25])
with self.assertRaisesWithPredicateMatch(
Exception, 'outcomes is not strictly increasing.'):
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
self.evaluate(dist.outcomes)
def testModeWithIntegerOutcomes(self):
outcomes = self._build_tensor([1, 2, 3])
probs = self._build_tensor([0.3, 0.1, 0.6])
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
mode = dist.mode()
self.assertAllEqual((), self._get_shape(mode))
self.assertAllEqual(3, mode)
def testRankOfOutcomesLargerThanOneRaises(self):
outcomes = self._build_tensor([[1.0, 2.0], [3.0, 4.0]])
probs = self._build_tensor([0.5, 0.5])
with self.assertRaisesWithPredicateMatch(
Exception, 'Rank of outcomes must be 1.'):
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
self.evaluate(dist.outcomes)
def testMean(self):
outcomes = self._build_tensor([1.0, 2.0])
probs = self._build_tensor([0.5, 0.5])
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
mean = dist.mean()
self.assertAllEqual((), self._get_shape(mean))
self.assertAllClose(1.5, mean)
def testCDFWithDifferentAtol(self):
outcomes = self._build_tensor([0.1, 0.2, 0.4, 0.8])
probs = self._build_tensor([0.0, 0.1, 0.2, 0.7])
x = self._build_tensor([[0.095, 0.095], [0.395, 0.395]])
dist1 = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
atol=0.001,
validate_args=True)
cdf = dist1.cdf(x)
self.assertAllEqual((2, 2), self._get_shape(cdf))
self.assertAllClose([[0.0, 0.0], [0.1, 0.1]], cdf)
dist2 = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
atol=0.01,
validate_args=True)
cdf = dist2.cdf(x)
self.assertAllEqual((2, 2), self._get_shape(cdf))
self.assertAllClose([[0.0, 0.0], [0.3, 0.3]], cdf)
def testSizeOfOutcomesIsZeroRaises(self):
outcomes = self._build_tensor([])
probs = self._build_tensor([])
with self.assertRaisesWithPredicateMatch(
Exception, 'Size of outcomes must be greater than 0.'):
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
self.evaluate(dist.outcomes)
def testMode(self):
outcomes = self._build_tensor([1.0, 2.0, 3.0])
probs = self._build_tensor([0.3, 0.1, 0.6])
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
mode = dist.mode()
self.assertAllEqual((), self._get_shape(mode))
self.assertAllClose(3.0, mode)
def testCDF(self):
outcomes = self._build_tensor([0.1, 0.2, 0.4, 0.8])
probs = self._build_tensor([0.0, 0.1, 0.2, 0.7])
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
cdf = dist.cdf(0.4)
self.assertAllEqual((), self._get_shape(cdf))
self.assertAllClose(0.3, cdf)
def testPMF(self):
outcomes = self._build_tensor([1.0, 2.0, 4.0, 8.0])
probs = self._build_tensor([[0.0, 0.1, 0.2, 0.7], [0.5, 0.3, 0.2,
0.0]])
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
prob = dist.prob(8.0)
self.assertAllEqual((2, ), self._get_shape(prob))
self.assertAllClose([0.7, 0.0], prob)
def testSample(self):
outcomes = self._build_tensor([1.0, 2.0])
probs = self._build_tensor([[0.2, 0.8], [0.8, 0.2]])
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
samples = self.evaluate(dist.sample(5000, seed=1234))
self.assertAllEqual((5000, 2), self._get_shape(samples))
self.assertAllClose(np.mean(samples, axis=0), dist.mean(), atol=0.1)
self.assertAllClose(np.std(samples, axis=0), dist.stddev(), atol=0.1)
def testEntropy(self):
outcomes = self._build_tensor([1, 2, 3, 4])
probs = np.array([[0.125, 0.125, 0.25, 0.5], [0.25, 0.25, 0.25, 0.25]])
outcome_probs = self._build_tensor(probs)
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=outcome_probs,
validate_args=True)
entropy = dist.entropy()
self.assertAllEqual((2, ), self._get_shape(entropy))
self.assertAllClose(np.sum(-probs * np.log(probs), axis=1), entropy)
def testMean(self):
outcomes = self._build_tensor([1.0, 2.0])
probs = self._build_tensor([[0.5, 0.5], [0.2, 0.8]])
expected_means = [1.5, 1.8]
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
mean = dist.mean()
self.assertAllEqual((2, ), self._get_shape(mean))
self.assertAllClose(expected_means, mean)
def testInequalLastDimRaises(self):
outcomes = self._build_tensor([1.0, 2.0])
probs = self._build_tensor([0.25, 0.25, 0.5])
with self.assertRaisesWithPredicateMatch(
Exception,
'Last dimension of outcomes and probs must be equal size'):
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
self.evaluate(dist.outcomes)
def testCDFWithBatchSampleShape(self):
outcomes = self._build_tensor([0.1, 0.2, 0.4, 0.8])
probs = self._build_tensor([0.0, 0.1, 0.2, 0.7])
x = self._build_tensor([[0.0999, 0.1], [0.2, 0.4], [0.8, 0.8001]])
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
cdf = dist.cdf(x)
self.assertAllEqual((3, 2), self._get_shape(cdf))
self.assertAllClose([[0.0, 0.0], [0.1, 0.3], [1.0, 1.0]], cdf)
def testPMFWithIntegerOutcomes(self):
outcomes = self._build_tensor([1, 2, 4, 8])
probs = self._build_tensor([0.0, 0.1, 0.2, 0.7])
x = self._build_tensor([[1], [2], [3], [4], [8]])
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
prob = dist.prob(x)
self.assertAllEqual((5, 1), self._get_shape(prob))
self.assertAllClose([[0.0], [0.1], [0.0], [0.2], [0.7]], prob)
def testPMFWithBatchSampleShape(self):
outcomes = self._build_tensor([1.0, 2.0, 4.0, 8.0])
probs = self._build_tensor([0.0, 0.1, 0.2, 0.7])
x = self._build_tensor([[1.0], [2.0], [3.0], [4.0], [8.0]])
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
prob = dist.prob(x)
self.assertAllEqual((5, 1), self._get_shape(prob))
self.assertAllClose([[0.0], [0.1], [0.0], [0.2], [0.7]], prob)
def testCDFWithIntegerOutcomes(self):
outcomes = self._build_tensor([1, 2, 4, 8])
probs = self._build_tensor([0.0, 0.1, 0.2, 0.7])
x = self._build_tensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
cdf = dist.cdf(x)
self.assertAllEqual((10, ), self._get_shape(cdf))
self.assertAllClose([0.0, 0.0, 0.1, 0.1, 0.3, 0.3, 0.3, 0.3, 1.0, 1.0],
cdf)
def testShape(self):
outcomes = self._build_tensor([0.0, 0.2, 0.3, 0.5])
logits = self._build_tensor([-0.1, 0.0, 0.1, 0.2])
dist = finite_discrete.FiniteDiscrete(outcomes,
logits=logits,
validate_args=True)
if self.use_static_shape:
self.assertAllEqual([], dist.batch_shape)
self.assertAllEqual([], dist.batch_shape_tensor())
self.assertAllEqual([], dist.event_shape)
self.assertAllEqual([], dist.event_shape_tensor())
def testStddevAndVariance(self):
outcomes = self._build_tensor([1.0, 2.0])
probs = self._build_tensor([[0.5, 0.5], [0.2, 0.8]])
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
stddev = dist.stddev()
self.assertAllEqual((2, ), self._get_shape(stddev))
self.assertAllClose([0.5, 0.4], stddev)
variance = dist.variance()
self.assertAllEqual((2, ), self._get_shape(variance))
self.assertAllClose([0.25, 0.16], variance)
def testParamTensorFromProbs(self):
outcomes = self._build_tensor([0.1, 0.2, 0.4])
x = tf.constant([0.1, 0.5, 0.4])
d = finite_discrete.FiniteDiscrete(outcomes,
probs=x,
validate_args=True)
self.assertAllClose(*self.evaluate(
[tf.math.log(x), d.logits_parameter()]),
atol=0,
rtol=1e-4)
self.assertAllClose(*self.evaluate([x, d.probs_parameter()]),
atol=0,
rtol=1e-4)
def testShapes(self):
outcomes = [0.0, 0.2, 0.3, 0.5]
outcomes_tensor = self._build_tensor(outcomes)
for batch_shape in ([1], [2], [3, 4, 5]):
logits = self._build_tensor(
np.random.uniform(-1,
1,
size=list(batch_shape) + [len(outcomes)]))
dist = finite_discrete.FiniteDiscrete(outcomes_tensor,
logits=logits,
validate_args=True)
if self.use_static_shape:
self.assertAllEqual(batch_shape, dist.batch_shape)
self.assertAllEqual(batch_shape, dist.batch_shape_tensor())
self.assertAllEqual([], dist.event_shape)
self.assertAllEqual([], dist.event_shape_tensor())
def testPMF(self):
outcomes = self._build_tensor([1.0, 2.0, 4.0, 8.0])
probs = self._build_tensor([0.0, 0.1, 0.2, 0.7])
dist = finite_discrete.FiniteDiscrete(outcomes,
probs=probs,
validate_args=True)
prob = dist.prob(4.0)
self.assertAllEqual((), self._get_shape(prob))
self.assertAllClose(0.2, prob)
# Outcome with zero probability.
prob = dist.prob(1.0)
self.assertAllEqual((), self._get_shape(prob))
self.assertAllClose(0.0, prob)
# Input that is not in the list of possible outcomes.
prob = dist.prob(3.0)
self.assertAllEqual((), self._get_shape(prob))
self.assertAllClose(0.0, prob)
