def testFloat32Fallback(self):
# Should be OK (float32 fallback).
self.polynomial = psd_kernels.Polynomial(bias_variance=0,
slope_variance=1,
exponent=1)
# Should be OK.
psd_kernels.Polynomial(bias_variance=np.float32(1.),
slope_variance=1.,
exponent=1.)
def testValidateArgsNonPositiveAreBad(self):
with self.assertRaisesOpError('Condition x > 0 did not hold'):
k = psd_kernels.Polynomial(bias_variance=-1., validate_args=True)
self.evaluate(k.bias_variance)
with self.assertRaisesOpError('Condition x > 0 did not hold'):
k = psd_kernels.Polynomial(slope_variance=-1., validate_args=True)
self.evaluate(k.slope_variance)
with self.assertRaisesOpError('Condition x > 0 did not hold'):
k = psd_kernels.Polynomial(exponent=-1., validate_args=True)
self.evaluate(k.exponent)
def testFloat32Fallback(self):
# Should be OK (float32 fallback).
self.polynomial = tfpk.Polynomial(
bias_variance=0,
slope_variance=1,
shift=0,
exponent=1)
# Should be OK.
tfpk.Polynomial(
bias_variance=np.float32(1.),
slope_variance=1.,
shift=0.,
exponent=1.)
def testValidateArgsNoneIsOk(self):
# No exception expected
k = psd_kernels.Polynomial(bias_variance=None,
slope_variance=None,
exponent=None,
validate_args=True)
self.evaluate(k.apply([[1.]], [[1.]]))
def testBatchShape(self, bias_variance, slope_variance, exponent, shape):
k = psd_kernels.Polynomial(bias_variance=bias_variance,
slope_variance=slope_variance,
exponent=exponent,
validate_args=True)
self.assertAllEqual(k.batch_shape.as_list(), shape)
self.assertAllEqual(self.evaluate(k.batch_shape_tensor()), shape)
def test_mismatched_float_types_are_bad(self):
with self.assertRaises(TypeError):
psd_kernels.Polynomial(
bias_variance=np.float32(1.),
slope_variance=np.float64(1.),
exponent=1.
)
def testValidateArgsNonPositiveAreBad(self):
with self.assertRaisesOpError('Condition x > 0 did not hold'):
k = tfpk.Polynomial(
bias_variance=-1.,
validate_args=True)
self.evaluate(k.apply([1.], [1.]))
with self.assertRaisesOpError('Condition x > 0 did not hold'):
k = tfpk.Polynomial(
slope_variance=-1.,
validate_args=True)
self.evaluate(k.apply([1.], [1.]))
with self.assertRaisesOpError('Condition x > 0 did not hold'):
k = tfpk.Polynomial(
exponent=-1.,
validate_args=True)
self.evaluate(k.apply([1.], [1.]))
def testShapesAreCorrectBroadcast(self):
k = psd_kernels.Polynomial(bias_variance=np.ones([2, 1, 1],
np.float32),
slope_variance=np.ones([1, 3, 1],
np.float32))
self.assertAllEqual(
k.matrix(np.ones([2, 4, 3], np.float32),
np.ones([2, 5, 3], np.float32)).shape, [2, 3, 2, 4, 5])
def testShapesAreCorrectMatrix(self, feature_ndims, x_shape, y_shape,
matrix_shape):
k = psd_kernels.Polynomial(bias_variance=0.,
slope_variance=1.,
exponent=1.,
feature_ndims=feature_ndims)
x = np.ones(x_shape, np.float32)
y = np.ones(y_shape, np.float32)
self.assertAllEqual(k.matrix(x, y).shape, matrix_shape)
def testFloat64(self):
# No exception expected
k = psd_kernels.Polynomial(bias_variance=np.float64(0.),
slope_variance=np.float64(1.),
exponent=np.float64(1.),
feature_ndims=1)
x = np.ones([5, 3], np.float64)
y = np.ones([5, 3], np.float64)
k.apply(x, y)
def testFloat32(self):
# No exception expected
k = psd_kernels.Polynomial(bias_variance=0.,
slope_variance=1.,
exponent=1.,
feature_ndims=1)
x = np.ones([5, 3], np.float32)
y = np.ones([5, 3], np.float32)
k.apply(x, y)
def testBatchShape(self, bias_variance, slope_variance,
shift, exponent, shape):
k = tfpk.Polynomial(
bias_variance=bias_variance,
slope_variance=slope_variance,
shift=shift,
exponent=exponent,
validate_args=True)
self.assertAllEqual(shape, k.batch_shape.as_list())
self.assertAllEqual(shape, self.evaluate(k.batch_shape_tensor()))
def testValuesAreCorrect(self):
bias_variance = 1.5
slope_variance = 0.5
exponent = 2
k = psd_kernels.Polynomial(bias_variance=bias_variance,
slope_variance=slope_variance,
exponent=exponent)
x = np.random.uniform(-1, 1, size=[5, 3]).astype(np.float32)
y = np.random.uniform(-1, 1, size=[4, 3]).astype(np.float32)
self.assertAllClose(
self.evaluate(k.matrix(x, y)),
bias_variance**2 + slope_variance**2 * (x.dot(y.T))**exponent)
def testShapesAreCorrectApply(self, feature_ndims,
x_shape, y_shape, apply_shape):
k = tfpk.Polynomial(
bias_variance=0.,
slope_variance=1.,
shift=0.,
exponent=1.,
feature_ndims=feature_ndims)
x = np.ones(x_shape, np.float32)
y = np.ones(y_shape, np.float32)
self.assertAllEqual(
apply_shape, k.apply(x, y).shape)
def testShapesAreCorrectBroadcast(self):
k = tfpk.Polynomial(
bias_variance=np.ones([2, 1, 1], np.float32),
slope_variance=np.ones([1, 3, 1], np.float32))
self.assertAllEqual(
[2, 3, 2, 4, 5],
#`--' | `--'
# | | `- matrix shape
# | `- from input batch shapes
# `- from broadcasting kernel params
k.matrix(
np.ones([2, 4, 3], np.float32),
np.ones([2, 5, 3], np.float32)
).shape)
def testValuesAreCorrect(self):
bias_variance = 1.5
slope_variance = 0.5
shift = 1.
exponent = 2
k = tfpk.Polynomial(
bias_variance=bias_variance,
slope_variance=slope_variance,
shift=shift,
exponent=exponent
)
x = np.random.uniform(-1, 1, size=[5, 3]).astype(np.float32)
y = np.random.uniform(-1, 1, size=[4, 3]).astype(np.float32)
self.assertAllClose(
(bias_variance ** 2 + slope_variance ** 2 *
((x - shift).dot((y - shift).T)) ** exponent),
self.evaluate(k.matrix(x, y))
)
def testNoneShapes(self):
k = tfpk.Polynomial(
bias_variance=np.reshape(np.arange(12.), [2, 3, 2]))
self.assertEqual([2, 3, 2], k.batch_shape.as_list())
def testShifttNonPositiveIsOk(self):
# No exception expected
k = tfpk.Polynomial(
shift=-1.,
validate_args=True)
self.evaluate(k.apply([1.], [1.]))
def testShifttNonPositiveIsOk(self):
# No exception expected
k = psd_kernels.Polynomial(shift=-1., validate_args=True)
self.evaluate(k.shift)
