def _fn(dtype, shape, name, trainable, add_variable_fn):
"""Creates a batch of `Deterministic` or `Normal` distributions."""
loc, scale = loc_scale_fn_(dtype, shape, name, trainable,
add_variable_fn)
if scale is None:
return deterministic_lib.Deterministic(loc=loc)
return normal_lib.Normal(loc=loc, scale=scale)
def testInvalidTolRaises(self):
loc = rng.rand(2, 3, 4).astype(np.float32)
with self.assertRaisesRegexp(errors.InvalidArgumentError,
"Condition x >= 0"):
_ = deterministic_lib.Deterministic(loc,
atol=-1,
validate_args=True)
def testSampleNoBatchDims(self):
deterministic = deterministic_lib.Deterministic(0.)
for sample_shape in [(), (4, )]:
with self.cached_session():
sample = deterministic.sample(sample_shape)
self.assertAllEqual(sample_shape, sample.get_shape())
self.assertAllClose(
np.zeros(sample_shape).astype(np.float32), sample.eval())
def testInvalidTolRaises(self):
loc = rng.rand(2, 3, 4).astype(np.float32)
deterministic = deterministic_lib.Deterministic(loc,
atol=-1,
validate_args=True)
with self.cached_session():
with self.assertRaisesOpError("Condition x >= 0"):
deterministic.prob(0.).eval()
def testProbWithDefaultTol(self):
loc = [[0., 1.], [2., 3.]]
x = [[0., 1.1], [1.99, 3.]]
deterministic = deterministic_lib.Deterministic(loc)
expected_prob = [[1., 0.], [0., 1.]]
with self.cached_session():
prob = deterministic.prob(x)
self.assertAllEqual((2, 2), prob.get_shape())
self.assertAllEqual(expected_prob, prob.eval())
def testCdfWithNonzeroRTol(self):
loc = [[1., 1.], [100., 100.]]
x = [[0.9, 1.], [99.9, 97]]
deterministic = deterministic_lib.Deterministic(loc, rtol=0.01)
expected_cdf = [[0., 1.], [1., 0.]]
with self.cached_session():
cdf = deterministic.cdf(x)
self.assertAllEqual((2, 2), cdf.get_shape())
self.assertAllEqual(expected_cdf, cdf.eval())
def testCdfWithNonzeroATol(self):
loc = [[0., 0.], [0., 0.]]
x = [[-1., -0.1], [-0.01, 1.000001]]
deterministic = deterministic_lib.Deterministic(loc, atol=0.05)
expected_cdf = [[0., 0.], [1., 1.]]
with self.cached_session():
cdf = deterministic.cdf(x)
self.assertAllEqual((2, 2), cdf.get_shape())
self.assertAllEqual(expected_cdf, cdf.eval())
def testProbWithNonzeroATolIntegerType(self):
loc = [[0, 1], [2, 3]]
x = [[0, 2], [4, 2]]
deterministic = deterministic_lib.Deterministic(loc, atol=1)
expected_prob = [[1, 1], [0, 1]]
with self.cached_session():
prob = deterministic.prob(x)
self.assertAllEqual((2, 2), prob.get_shape())
self.assertAllEqual(expected_prob, prob.eval())
def testProbWithNonzeroATol(self):
loc = [[0., 1.], [2., 3.]]
x = [[0., 1.1], [1.99, 3.]]
deterministic = deterministic_lib.Deterministic(loc, atol=0.05)
expected_prob = [[1., 0.], [1., 1.]]
with self.test_session():
prob = deterministic.prob(x)
self.assertAllEqual((2, 2), prob.get_shape())
self.assertAllEqual(expected_prob, prob.eval())
def testSampleWithBatchDims(self):
deterministic = deterministic_lib.Deterministic([0., 0.])
for sample_shape in [(), (4, )]:
with self.test_session():
sample = deterministic.sample(sample_shape)
self.assertAllEqual(sample_shape + (2, ), sample.get_shape())
self.assertAllClose(
np.zeros(sample_shape + (2, )).astype(np.float32),
sample.eval())
def testProbWithNonzeroRTol(self):
loc = [[0., 1.], [100., 100.]]
x = [[0., 1.1], [100.1, 103.]]
deterministic = deterministic_lib.Deterministic(loc, rtol=0.01)
expected_prob = [[1., 0.], [1., 0.]]
with self.cached_session():
prob = deterministic.prob(x)
self.assertAllEqual((2, 2), prob.get_shape())
self.assertAllEqual(expected_prob, prob.eval())
def _fn(dtype, shape, name, trainable, add_variable_fn):
"""Creates multivariate `Deterministic` or `Normal` distribution."""
loc, scale = loc_scale_fn_(dtype, shape, name, trainable, add_variable_fn)
if scale is None:
dist = deterministic_lib.Deterministic(loc=loc)
else:
dist = normal_lib.Normal(loc=loc, scale=scale)
reinterpreted_batch_ndims = array_ops.shape(dist.batch_shape_tensor())[0]
return independent_lib.Independent(
dist, reinterpreted_batch_ndims=reinterpreted_batch_ndims)
def testProbWithNonzeroRTolIntegerType(self):
loc = [[10, 10, 10], [10, 10, 10]]
x = [[10, 20, 30], [10, 20, 30]]
# Batch 0 will have rtol = 0
# Batch 1 will have rtol = 1 (100% slack allowed)
deterministic = deterministic_lib.Deterministic(loc, rtol=[[0], [1]])
expected_prob = [[1, 0, 0], [1, 1, 0]]
with self.cached_session():
prob = deterministic.prob(x)
self.assertAllEqual((2, 3), prob.get_shape())
self.assertAllEqual(expected_prob, prob.eval())
def testShape(self):
with self.cached_session():
loc = rng.rand(2, 3, 4)
deterministic = deterministic_lib.Deterministic(loc)
self.assertAllEqual(deterministic.batch_shape_tensor().eval(),
(2, 3, 4))
self.assertAllEqual(deterministic.batch_shape, (2, 3, 4))
self.assertAllEqual(deterministic.event_shape_tensor().eval(), [])
self.assertEqual(deterministic.event_shape,
tensor_shape.TensorShape([]))
def testSampleDynamicWithBatchDims(self):
loc = array_ops.placeholder(np.float32)
sample_shape = array_ops.placeholder(np.int32)
deterministic = deterministic_lib.Deterministic(loc)
for sample_shape_ in [(), (4, )]:
with self.cached_session():
sample_ = deterministic.sample(sample_shape).eval(
feed_dict={
loc: [0., 0.],
sample_shape: sample_shape_
})
self.assertAllClose(
np.zeros(sample_shape_ + (2, )).astype(np.float32),
sample_)
def testEntropy(self):
loc = np.array([-0.1, -3.2, 7.])
deterministic = deterministic_lib.Deterministic(loc=loc)
with self.cached_session() as sess:
entropy_ = sess.run(deterministic.entropy())
self.assertAllEqual(np.zeros(3), entropy_)
def testProbWithNoBatchDims(self):
deterministic = deterministic_lib.Deterministic(0.)
with self.cached_session():
self.assertAllClose(1., deterministic.prob(0.).eval())
self.assertAllClose(0., deterministic.prob(2.).eval())
self.assertAllClose([1., 0.], deterministic.prob([0., 2.]).eval())
def testProbWithNoBatchDimsIntegerType(self):
deterministic = deterministic_lib.Deterministic(0)
with self.cached_session():
self.assertAllClose(1, deterministic.prob(0).eval())
self.assertAllClose(0, deterministic.prob(2).eval())
self.assertAllClose([1, 0], deterministic.prob([0, 2]).eval())
