def __eq__(self, other):
return B.diag(self) == B.diag(other)
def dim(self):
"""Dimensionality."""
return B.shape(self.var)[0]
def lmatmul(self, other):
return Normal(B.dot(B.dot(other, self.var), other, tr_b=True),
B.dot(other, self.mean))
def __init__(self, z, e, x, y, ref=None):
Observations.__init__(self, x, y, ref=ref)
# Extract processes.
p_x, x = type_parameter(self.x), self.x.get()
z = ensure_at(z, self._ref)
p_z, z = type_parameter(z), z.get()
# Construct the necessary kernel matrices.
K_zx = self.graph.kernels[p_z, p_x](z, x)
K_z = self.graph.kernels[p_z](z)
# Evaluating `e.kernel(x)` will yield incorrect results if `x` is a
# `MultiInput`, because `x` then still designates the particular
# components of `f`. Fix that by instead designating the elements of
# `e`.
if isinstance(x, MultiInput):
x_n = MultiInput(*(p(xi.get())
for p, xi in zip(e.kernel.ps, x.get())))
else:
x_n = x
# Construct the noise kernel matrix.
K_n = e.kernel(x_n)
# The approximation can only handle diagonal noise matrices.
if not isinstance(K_n, Diagonal):
raise RuntimeError('Kernel matrix of noise must be diagonal.')
# And construct the components for the inducing point approximation.
L_z = B.cholesky(matrix(K_z))
A = B.eye_from(K_z) + B.qf(K_n, B.transpose(B.trisolve(L_z, K_zx)))
y_bar = uprank(self.y) - e.mean(x_n) - self.graph.means[p_x](x)
prod_y_bar = B.trisolve(L_z, B.qf(K_n, B.transpose(K_zx), y_bar))
# Compute the optimal mean.
mean = self.graph.means[p_z](z) + \
B.qf(A, B.trisolve(L_z, K_z), prod_y_bar)
# Compute the ELBO.
# NOTE: The calculation of `trace_part` asserts that `K_n` is diagonal.
# The rest, however, is completely generic.
trace_part = B.ratio(
Diagonal(self.graph.kernels[p_x].elwise(x)[:, 0]) -
Diagonal(B.qf_diag(K_z, K_zx)), K_n)
det_part = B.logdet(2 * B.pi * K_n) + B.logdet(A)
qf_part = B.qf(K_n, y_bar)[0, 0] - B.qf(A, prod_y_bar)[0, 0]
elbo = -0.5 * (trace_part + det_part + qf_part)
# Store relevant quantities.
self.elbo = elbo
self.A = A
# Update observations to reflect pseudo-points.
self.x = p_z(z)
self.y = mean
def __eq__(self, other):
return B.all(self.scale == other.scale) and self[0] == other[0]
def test_qf():
# Generate some test inputs.
b, c = np.random.randn(5, 3), np.random.randn(5, 3)
# Generate some matrices to test.
a = np.random.randn(5, 5)
a = Dense(a.dot(a.T))
d = Diagonal(B.diag(a))
e = np.random.randn(2, 2)
wb = d + LowRank(left=np.random.randn(5, 2),
middle=e.dot(e.T))
for x in [a, d, wb]:
allclose(B.qf(x, b), np.linalg.solve(to_np(x), b).T.dot(b))
allclose(B.qf(x, b, b), B.qf(x, b))
allclose(B.qf(x, b, c), np.linalg.solve(to_np(x), b).T.dot(c))
allclose(B.qf_diag(x, b),
np.diag(np.linalg.solve(to_np(x), b).T.dot(b)))
allclose(B.qf_diag(x, b, b), B.qf_diag(x, b, b))
allclose(B.qf_diag(x, b, c),
np.diag(np.linalg.solve(to_np(x), b).T.dot(c)))
# Test `LowRank`.
lr = LowRank(np.random.randn(5, 3))
with pytest.raises(RuntimeError):
B.qf(lr, b)
with pytest.raises(RuntimeError):
B.qf(lr, b, c)
with pytest.raises(RuntimeError):
B.qf_diag(lr, b)
with pytest.raises(RuntimeError):
B.qf_diag(lr, b, c)
def test_shorthands():
a = Dense(np.random.randn(4, 4))
allclose(a.T, B.transpose(a))
allclose(a.__matmul__(a), B.matmul(a, a))
def __call__(self, x):
return B.concat(*[self(xi) for xi in x.get()], axis=0)
def _take_x(k: MultiOutputKernel, x: FDD, mask: B.Numeric):
if x.p not in k.ps:
raise ValueError(f"Process {x.p} is not part of the multi-output kernel.")
return B.take(x, mask)
def elwise(self, x, y):
if len(x.get()) != len(y.get()):
raise ValueError('MultiOutputKernel.elwise must be called with '
'similarly sized MultiInputs.')
return B.concat(*[self.elwise(xi, yi)
for xi, yi in zip(x.get(), y.get())], axis=0)
def to_np(a: AbstractMatrix):
return B.dense(a)
def __call__(self, x, y):
return B.block(*[[self(xi, yi) for yi in y.get()] for xi in x.get()])
def qf_diag(a, b, c):
return B.diag(B.qf(a, b, c))
def qf_diag(a, b):
return B.qf_diag(a, b, b)
def test_inverse_and_logdet():
# Test `Dense`.
a = np.random.randn(3, 3)
a = Dense(a.dot(a.T))
allclose(B.matmul(a, B.inverse(a)), np.eye(3))
allclose(B.matmul(B.inverse(a), a), np.eye(3))
allclose(B.logdet(a), np.log(np.linalg.det(to_np(a))))
# Test `Diagonal`.
d = Diagonal(np.array([1, 2, 3]))
allclose(B.matmul(d, B.inverse(d)), np.eye(3))
allclose(B.matmul(B.inverse(d), d), np.eye(3))
allclose(B.logdet(d), np.log(np.linalg.det(to_np(d))))
assert B.shape(B.inverse(Diagonal(np.array([1, 2]),
rows=2, cols=4))) == (4, 2)
# Test `Woodbury`.
a = np.random.randn(3, 2)
b = np.random.randn(2, 2) + 1e-2 * np.eye(2)
wb = d + LowRank(left=a, middle=b.dot(b.T))
for _ in range(4):
allclose(B.matmul(wb, B.inverse(wb)), np.eye(3))
allclose(B.matmul(B.inverse(wb), wb), np.eye(3))
allclose(B.logdet(wb), np.log(np.linalg.det(to_np(wb))))
wb = B.inverse(wb)
# Test `LowRank`.
with pytest.raises(RuntimeError):
B.inverse(wb.lr)
with pytest.raises(RuntimeError):
B.logdet(wb.lr)
def test_normal_1d():
# Test broadcasting.
d = Normal1D(1, 0)
assert type(d.var) == UniformlyDiagonal
assert B.shape(d.var) == (1, 1)
assert B.shape(d.mean) == (1, 1)
d = Normal1D(1, np.array([0, 0, 0]))
assert type(d.var) == UniformlyDiagonal
assert B.shape(d.var) == (3, 3)
assert B.shape(d.mean) == (3, 1)
d = Normal1D(np.array([1, 2, 3]), 0)
assert type(d.var) == Diagonal
assert B.shape(d.var) == (3, 3)
assert B.shape(d.mean) == (3, 1)
d = Normal1D(np.array([1, 2, 3]), np.array([0, 0, 0]))
assert type(d.var) == Diagonal
assert B.shape(d.var) == (3, 3)
assert B.shape(d.mean) == (3, 1)
d = Normal1D(1)
assert type(d.var) == UniformlyDiagonal
assert B.shape(d.var) == (1, 1)
assert B.shape(d.mean) == (1, 1)
d = Normal1D(np.array([1, 2, 3]))
assert type(d.var) == Diagonal
assert B.shape(d.var) == (3, 3)
assert B.shape(d.mean) == (3, 1)
with pytest.raises(ValueError):
Normal1D(np.eye(3))
with pytest.raises(ValueError):
Normal1D(np.eye(3), 0)
with pytest.raises(ValueError):
Normal1D(1, np.ones((3, 1)))
with pytest.raises(ValueError):
Normal1D(np.array([1, 2]), np.ones((3, 1)))
def test_lr_diff():
# First, test correctness.
a = np.random.randn(3, 2)
b = np.random.randn(2, 2)
lr1 = LowRank(left=a, right=np.random.randn(3, 2), middle=b.dot(b.T))
a = np.random.randn(3, 2)
b = np.random.randn(2, 2)
lr2 = LowRank(left=a, right=np.random.randn(3, 2), middle=b.dot(b.T))
allclose(B.lr_diff(lr1, lr1), B.zeros(3, 3))
allclose(B.lr_diff(lr1 + lr2, lr1), lr2)
allclose(B.lr_diff(lr1 + lr2, lr2), lr1)
allclose(B.lr_diff(lr1 + lr1 + lr2, lr1), lr1 + lr2)
allclose(B.lr_diff(lr1 + lr1 + lr2, lr2), lr1 + lr1)
allclose(B.lr_diff(lr1 + lr1 + lr2, lr1 + lr1), lr2)
allclose(B.lr_diff(lr1 + lr1 + lr2, lr1 + lr2), lr1)
allclose(B.lr_diff(lr1 + lr1 + lr2, lr1 + lr1 + lr2), B.zeros(3, 3))
# Second, test positive definiteness.
lr1 = LowRank(left=lr1.left, middle=lr1.middle)
lr2 = LowRank(left=lr2.left, middle=lr2.middle)
B.cholesky(B.lr_diff(lr1, 0.999 * lr1))
B.cholesky(B.lr_diff(lr1 + lr2, lr1))
B.cholesky(B.lr_diff(lr1 + lr2, lr2))
B.cholesky(B.lr_diff(lr1 + lr1 + lr2, lr1))
B.cholesky(B.lr_diff(lr1 + lr1 + lr2, lr1 + lr1))
B.cholesky(B.lr_diff(lr1 + lr1 + lr2, lr1 + lr2))
B.cholesky(B.lr_diff(lr1 + lr1 + lr2, lr1 + lr1 + 0.999 * lr2))
def test_uprank():
allclose(uprank(0), [[0]])
allclose(uprank(np.array([0])), [[0]])
allclose(uprank(np.array([[0]])), [[0]])
assert type(uprank(Component('test')(0))) == Component('test')
k = OneKernel()
assert B.shape(k(0, 0)) == (1, 1)
assert B.shape(k(0, np.ones(5))) == (1, 5)
assert B.shape(k(0, np.ones((5, 2)))) == (1, 5)
assert B.shape(k(np.ones(5), 0)) == (5, 1)
assert B.shape(k(np.ones(5), np.ones(5))) == (5, 5)
assert B.shape(k(np.ones(5), np.ones((5, 2)))) == (5, 5)
assert B.shape(k(np.ones((5, 2)), 0)) == (5, 1)
assert B.shape(k(np.ones((5, 2)), np.ones(5))) == (5, 5)
assert B.shape(k(np.ones((5, 2)), np.ones((5, 2)))) == (5, 5)
with pytest.raises(ValueError):
k(0, np.ones((5, 2, 1)))
with pytest.raises(ValueError):
k(np.ones((5, 2, 1)))
m = OneMean()
assert B.shape(m(0)) == (1, 1)
assert B.shape(m(np.ones(5))) == (5, 1)
assert B.shape(m(np.ones((5, 2)))) == (5, 1)
p = GP(EQ(), graph=Graph())
x = np.linspace(0, 10, 10)
approx(p.condition(1, 1)(1).mean, np.array([[1]]))
approx(p.condition(x, x)(x).mean, x[:, None])
approx(p.condition(x, x[:, None])(x).mean, x[:, None])
def compare(a):
allclose(B.transpose(a), to_np(a).T)
def test_block_matrix():
dt = np.float64
# Check correctness.
rows = [[np.random.randn(4, 3), np.random.randn(4, 5)],
[np.random.randn(6, 3), np.random.randn(6, 5)]]
allclose(B.block_matrix(*rows), B.concat2d(*rows))
# Check that grid is checked correctly.
assert type(B.block_matrix([Zero(dt, 3, 7), Zero(dt, 3, 4)],
[Zero(dt, 4, 5), Zero(dt, 4, 6)])) == Dense
with pytest.raises(ValueError):
B.block_matrix([Zero(dt, 5, 5), Zero(dt, 3, 6)],
[Zero(dt, 2, 5), Zero(dt, 4, 6)])
# Test zeros.
res = B.block_matrix([Zero(dt, 3, 5), Zero(dt, 3, 6)],
[Zero(dt, 4, 5), Zero(dt, 4, 6)])
assert type(res) == Zero
allclose(res, Zero(dt, 7, 11))
# Test ones.
res = B.block_matrix([One(dt, 3, 5), One(dt, 3, 6)],
[One(dt, 4, 5), One(dt, 4, 6)])
assert type(res) == One
allclose(res, One(dt, 7, 11))
# Test diagonal.
res = B.block_matrix([Diagonal(np.array([1, 2])), Zero(dt, 2, 3)],
[Zero(dt, 3, 2), Diagonal(np.array([3, 4, 5]))])
assert type(res) == Diagonal
allclose(res, Diagonal(np.array([1, 2, 3, 4, 5])))
# Check that all blocks on the diagonal must be diagonal or zero.
assert type(B.block_matrix([Diagonal(np.array([1, 2])), Zero(dt, 2, 3)],
[Zero(dt, 3, 2), One(dt, 3)])) == Dense
assert type(B.block_matrix([Diagonal(np.array([1, 2])), Zero(dt, 2, 3)],
[Zero(dt, 3, 2), Zero(dt, 3)])) == Diagonal
# Check that all blocks on the diagonal must be square.
assert type(B.block_matrix([Diagonal(np.array([1, 2])), Zero(dt, 2, 4)],
[Zero(dt, 3, 2), Zero(dt, 3, 4)])) == Dense
# Check that all other blocks must be zero.
assert type(B.block_matrix([Diagonal(np.array([1, 2])), One(dt, 2, 3)],
[Zero(dt, 3, 2),
Diagonal(np.array([3, 4, 5]))])) == Dense
def test_eye():
a = Dense(np.random.randn(5, 10))
allclose(B.eye(a), np.eye(5, 10))
allclose(B.eye(a.T), np.eye(10, 5))
def test_trace():
a = np.random.randn(10, 10)
allclose(B.trace(Dense(a)), np.trace(a))
def ones(x):
return B.ones([B.shape(x)[0], 1], dtype=B.dtype(x))
def test_diag_len():
assert B.diag_len(np.ones((5, 5))) == 5
assert B.diag_len(np.ones((10, 5))) == 5
assert B.diag_len(np.ones((5, 10))) == 5
def dtype(self):
"""Data type."""
return B.dtype(self.var)
def test_matmul():
diag_square = Diagonal(np.array([1, 2]), 3)
diag_tall = Diagonal(np.array([3, 4]), 5, 3)
diag_wide = Diagonal(np.array([5, 6]), 2, 3)
dense_square = Dense(np.random.randn(3, 3))
dense_tall = Dense(np.random.randn(5, 3))
dense_wide = Dense(np.random.randn(2, 3))
lr = LowRank(left=np.random.randn(5, 2),
right=np.random.randn(3, 2),
middle=np.random.randn(2, 2))
def compare(a, b):
return np.allclose(to_np(B.matmul(a, b)),
B.matmul(to_np(a), to_np(b)))
# Test `Dense`.
assert compare(dense_wide, dense_tall.T), 'dense w x dense t'
# Test `LowRank`.
assert compare(lr, dense_tall.T), 'lr x dense t'
assert compare(dense_wide, lr.T), 'dense w x lr'
assert compare(lr, diag_tall.T), 'lr x diag t'
assert compare(diag_wide, lr.T), 'diag w x lr'
assert compare(lr, lr.T), 'lr x lr'
assert compare(lr.T, lr), 'lr x lr (2)'
# Test `Diagonal`.
# Test multiplication between diagonal matrices.
assert compare(diag_square, diag_square.T), 'diag s x diag s'
assert compare(diag_tall, diag_square.T), 'diag t x diag s'
assert compare(diag_wide, diag_square.T), 'diag w x diag s'
assert compare(diag_square, diag_tall.T), 'diag s x diag t'
assert compare(diag_tall, diag_tall.T), 'diag t x diag t'
assert compare(diag_wide, diag_tall.T), 'diag w x diag t'
assert compare(diag_square, diag_wide.T), 'diag s x diag w'
assert compare(diag_tall, diag_wide.T), 'diag t x diag w'
assert compare(diag_wide, diag_wide.T), 'diag w x diag w'
# Test multiplication between diagonal and dense matrices.
assert compare(diag_square, dense_square.T), 'diag s x dense s'
assert compare(diag_square, dense_tall.T), 'diag s x dense t'
assert compare(diag_square, dense_wide.T), 'diag s x dense w'
assert compare(diag_tall, dense_square.T), 'diag t x dense s'
assert compare(diag_tall, dense_tall.T), 'diag t x dense t'
assert compare(diag_tall, dense_wide.T), 'diag t x dense w'
assert compare(diag_wide, dense_square.T), 'diag w x dense s'
assert compare(diag_wide, dense_tall.T), 'diag w x dense t'
assert compare(diag_wide, dense_wide.T), 'diag w x dense w'
assert compare(dense_square, diag_square.T), 'dense s x diag s'
assert compare(dense_square, diag_tall.T), 'dense s x diag t'
assert compare(dense_square, diag_wide.T), 'dense s x diag w'
assert compare(dense_tall, diag_square.T), 'dense t x diag s'
assert compare(dense_tall, diag_tall.T), 'dense t x diag t'
assert compare(dense_tall, diag_wide.T), 'dense t x diag w'
assert compare(dense_wide, diag_square.T), 'dense w x diag s'
assert compare(dense_wide, diag_tall.T), 'dense w x diag t'
assert compare(dense_wide, diag_wide.T), 'dense w x diag w'
# Test `B.matmul` with three matrices simultaneously.
allclose(B.matmul(dense_tall, dense_square, dense_wide, tr_c=True),
(to_np(dense_tall)
.dot(to_np(dense_square))
.dot(to_np(dense_wide).T)))
# Test `Woodbury`.
wb = lr + dense_tall
assert compare(wb, dense_square.T), 'wb x dense s'
assert compare(dense_square, wb.T), 'dense s x wb'
assert compare(wb, wb.T), 'wb x wb'
assert compare(wb.T, wb), 'wb x wb (2)'
def m2(self):
"""Second moment."""
return self.var + B.outer(B.squeeze(self.mean))
def compare(a, b):
return np.allclose(to_np(B.matmul(a, b)),
B.matmul(to_np(a), to_np(b)))
def __init__(self, var, mean=None):
# Consider all various ranks of `var` and `mean` for convenient
# broadcasting behaviour.
if mean is not None:
if B.rank(var) == 1:
if B.rank(mean) == 0:
mean = mean * B.ones(B.dtype(var), B.shape(var)[0], 1)
elif B.rank(mean) == 1:
mean = mean[:, None]
else:
raise ValueError('Invalid rank {} of mean.'
''.format(B.rank(mean)))
var = Diagonal(var)
elif B.rank(var) == 0:
if B.rank(mean) == 0:
mean = mean * B.ones(B.dtype(var), 1, 1)
var = UniformlyDiagonal(var, 1)
elif B.rank(mean) == 1:
mean = mean[:, None]
var = UniformlyDiagonal(var, B.shape(mean)[0])
else:
raise ValueError('Invalid rank {} of mean.'
''.format(B.rank(mean)))
else:
raise ValueError('Invalid rank {} of variance.'
''.format(B.rank(var)))
else:
if B.rank(var) == 0:
var = UniformlyDiagonal(var, 1)
elif B.rank(var) == 1:
var = Diagonal(var)
else:
raise ValueError('Invalid rank {} of variance.'
''.format(B.rank(var)))
Normal.__init__(self, var, mean)
def qf(a, b, c):
return B.matmul(b, B.inverse(a), c, tr_a=True)
