def test_frozen_matrix_normal(self):
for i in range(1, 5):
for j in range(1, 5):
M = 0.3 * np.ones((i, j))
U = 0.5 * np.identity(i) + 0.5 * np.ones((i, i))
V = 0.7 * np.identity(j) + 0.3 * np.ones((j, j))
frozen = matrix_normal(mean=M, rowcov=U, colcov=V)
rvs1 = frozen.rvs(random_state=1234)
rvs2 = matrix_normal.rvs(mean=M,
rowcov=U,
colcov=V,
random_state=1234)
assert_equal(rvs1, rvs2)
X = frozen.rvs(random_state=1234)
pdf1 = frozen.pdf(X)
pdf2 = matrix_normal.pdf(X, mean=M, rowcov=U, colcov=V)
assert_equal(pdf1, pdf2)
logpdf1 = frozen.logpdf(X)
logpdf2 = matrix_normal.logpdf(X, mean=M, rowcov=U, colcov=V)
assert_equal(logpdf1, logpdf2)
def test_array_input(self):
# Check array of inputs has the same output as the separate entries.
num_rows = 4
num_cols = 3
M = 0.3 * np.ones((num_rows, num_cols))
U = 0.5 * np.identity(num_rows) + 0.5 * np.ones((num_rows, num_rows))
V = 0.7 * np.identity(num_cols) + 0.3 * np.ones((num_cols, num_cols))
N = 10
frozen = matrix_normal(mean=M, rowcov=U, colcov=V)
X1 = frozen.rvs(size=N, random_state=1234)
X2 = frozen.rvs(size=N, random_state=4321)
X = np.concatenate((X1[np.newaxis, :, :, :], X2[np.newaxis, :, :, :]),
axis=0)
assert_equal(X.shape, (2, N, num_rows, num_cols))
array_logpdf = frozen.logpdf(X)
assert_equal(array_logpdf.shape, (2, N))
for i in range(2):
for j in range(N):
separate_logpdf = matrix_normal.logpdf(X[i, j],
mean=M,
rowcov=U,
colcov=V)
assert_allclose(separate_logpdf, array_logpdf[i, j], 1E-10)
def testLogPDFScalarBatch(self):
seed_stream = test_util.test_seed_stream()
loc, scale_row, scale_col = self._random_loc_and_scale([], [2, 3],
seed_stream)
matrix_normal = tfd.MatrixNormalLinearOperator(loc,
scale_row,
scale_col,
validate_args=True)
x = tf.random.normal(shape=[2, 3],
seed=seed_stream(),
dtype=self.dtype)
log_pdf = matrix_normal.log_prob(x)
pdf = matrix_normal.prob(x)
loc_, row_cov_, col_cov_, log_pdf_, pdf_, x_ = self.evaluate([
loc,
tf.linalg.matmul(scale_row.to_dense(),
scale_row.to_dense(),
adjoint_b=True),
tf.linalg.matmul(scale_col.to_dense(),
scale_col.to_dense(),
adjoint_b=True), log_pdf, pdf, x
])
scipy_matrix_normal = stats.matrix_normal(mean=loc_,
rowcov=row_cov_,
colcov=col_cov_)
self.assertEqual((), log_pdf_.shape)
self.assertEqual((), pdf_.shape)
self.assertAllClose(scipy_matrix_normal.logpdf(x_), log_pdf_)
self.assertAllClose(scipy_matrix_normal.pdf(x_), pdf_)
def testLogPDFBatch(self):
seed_stream = test_util.test_seed_stream()
loc, scale_row, scale_col = self._random_loc_and_scale(
batch_shape=[3], matrix_shape=[3, 5], seed_stream=seed_stream)
matrix_normal = tfd.MatrixNormalLinearOperator(loc,
scale_row,
scale_col,
validate_args=True)
x = tf.random.normal(shape=[3, 3, 5],
seed=seed_stream(),
dtype=self.dtype)
log_pdf = matrix_normal.log_prob(x)
pdf = matrix_normal.prob(x)
loc_, row_cov_, col_cov_, log_pdf_, pdf_, x_ = self.evaluate([
loc,
tf.linalg.matmul(scale_row.to_dense(),
scale_row.to_dense(),
adjoint_b=True),
tf.linalg.matmul(scale_col.to_dense(),
scale_col.to_dense(),
adjoint_b=True), log_pdf, pdf, x
])
self.assertEqual((3, ), log_pdf_.shape)
self.assertEqual((3, ), pdf_.shape)
for i in range(3):
scipy_matrix_normal = stats.matrix_normal(mean=loc_[i],
rowcov=row_cov_[i],
colcov=col_cov_[i])
expected_log_pdf = scipy_matrix_normal.logpdf(x_[i])
expected_pdf = scipy_matrix_normal.pdf(x_[i])
self.assertAllClose(expected_log_pdf, log_pdf_[i], rtol=3e-5)
self.assertAllClose(expected_pdf, pdf_[i], rtol=3e-5)
def sample_theta(self, data_statistics):
"""
:param data_statistics: output of self.compute_statistics()
:return:
"""
S_xx, S_yy, S_yx = data_statistics
for k in range(self.L):
s_xx = S_xx[:, :, k] + self.sampling_parameters['K']
s_yx = S_yx[:, :, k] + self.sampling_parameters[
'M'] @ self.sampling_parameters['K']
s_yy = S_yy[:, :, k] + self.sampling_parameters['M'] @ (
self.sampling_parameters['K']
@ self.sampling_parameters['M'].T)
M = s_yx @ inv(s_xx)
s_ygx = s_yy - M @ s_yx.T
# enforce symmetry
s_ygx = (s_ygx + s_ygx.T) / 2
sigma = invwishart(df=self.sampling_parameters['n_0'] +
self.state_counts['Ns'][k],
scale=self.sampling_parameters['S_0'] +
s_ygx).rvs()
A = matrix_normal(mean=M + self.sampling_parameters['M'],
rowcov=sigma,
colcov=inv(s_xx)).rvs()
self.theta['A'][:, :, k] = A
self.theta['inv_sigma'][:, :, k] = inv(sigma)
def generate_expressionTrait(self, MatrixVariate=False):
self.generate_parameters()
term1 = self.X @ self.beta
term2 = self.G @ self.delta
if MatrixVariate:
try:
U = matrix_normal(rowcov=self.K, colcov=self.rho_u).rvs(1)
except:
K = self.K + 1e-8 * np.eye(self.num_sample)
U = matrix_normal(rowcov=K, colcov=self.rho_u).rvs(1)
epsilon = matrix_normal(rowcov=np.diag(np.ones(self.num_sample)),
colcov=self.rho_e).rvs(1)
term3 = U @ np.sqrt(self.xi)
term4 = epsilon @ np.sqrt(
np.diag(np.ones(self.num_trait)) - self.xi)
term5 = (term3 + term4) @ np.sqrt(self.eta)
self.Y = term1 + term2 + term5
else:
Vu = np.sqrt(self.eta) @ np.sqrt(self.xi) @ self.rho_u @ np.sqrt(
self.xi) @ np.sqrt(self.eta)
Ve = np.sqrt(self.eta) @ np.sqrt(np.eye(self.num_trait) -
self.xi) @ self.rho_u @ np.sqrt(
np.eye(self.num_trait) -
self.xi) @ np.sqrt(self.eta)
mean = np.zeros(self.num_trait)
U, Lambda, _ = np.linalg.svd(self.K)
Lambda[Lambda < 1e-13] = 1e-13
_vec_e = np.zeros([self.num_trait, self.num_sample])
now = datetime.now().time()
print('******************************* Generating Y:',
" now the time is ", now)
print(self.num_sample)
for n in range(self.num_sample):
if (n + 1) % 20 == 0:
print('sampling : {} of {}'.format(n + 1, self.num_sample),
flush=True)
cov = Lambda[n] * Vu + Ve
_vec_e[:, n] = np.random.multivariate_normal(mean=mean,
cov=cov,
size=1)
vec_e = U @ _vec_e.T
self.Y = term1 + term2 + vec_e
def sample_init_theta(self, seed=None):
for k in range(self.L):
sigma = invwishart(df=self.sampling_parameters['n_0'] +
self.state_counts['Ns'][k],
scale=self.sampling_parameters['S_0'] + 0,
seed=seed).rvs()
A = matrix_normal(mean=self.sampling_parameters['M'],
rowcov=sigma,
colcov=inv(self.sampling_parameters['K'])).rvs()
self.theta['A'][:, :, k] = A
self.theta['inv_sigma'][:, :, k] = inv(sigma)
def test_covariance_expansion(self):
# Check that covariance can be specified with scalar or vector
num_rows = 4
num_cols = 3
M = 0.3 * np.ones((num_rows,num_cols))
Uv = 0.2*np.ones(num_rows)
Us = 0.2
Vv = 0.1*np.ones(num_cols)
Vs = 0.1
Ir = np.identity(num_rows)
Ic = np.identity(num_cols)
assert_equal(matrix_normal(mean=M, rowcov=Uv, colcov=Vv).rowcov,
0.2*Ir)
assert_equal(matrix_normal(mean=M, rowcov=Uv, colcov=Vv).colcov,
0.1*Ic)
assert_equal(matrix_normal(mean=M, rowcov=Us, colcov=Vs).rowcov,
0.2*Ir)
assert_equal(matrix_normal(mean=M, rowcov=Us, colcov=Vs).colcov,
0.1*Ic)
def test_covariance_expansion(self):
# Check that covariance can be specified with scalar or vector
num_rows = 4
num_cols = 3
M = 0.3 * np.ones((num_rows, num_cols))
Uv = 0.2 * np.ones(num_rows)
Us = 0.2
Vv = 0.1 * np.ones(num_cols)
Vs = 0.1
Ir = np.identity(num_rows)
Ic = np.identity(num_cols)
assert_equal(
matrix_normal(mean=M, rowcov=Uv, colcov=Vv).rowcov, 0.2 * Ir)
assert_equal(
matrix_normal(mean=M, rowcov=Uv, colcov=Vv).colcov, 0.1 * Ic)
assert_equal(
matrix_normal(mean=M, rowcov=Us, colcov=Vs).rowcov, 0.2 * Ir)
assert_equal(
matrix_normal(mean=M, rowcov=Us, colcov=Vs).colcov, 0.1 * Ic)
def test_default_inputs(self):
# Check that default argument handling works
num_rows = 4
num_cols = 3
M = 0.3 * np.ones((num_rows,num_cols))
U = 0.5 * np.identity(num_rows) + 0.5 * np.ones((num_rows, num_rows))
V = 0.7 * np.identity(num_cols) + 0.3 * np.ones((num_cols, num_cols))
Z = np.zeros((num_rows, num_cols))
Zr = np.zeros((num_rows, 1))
Zc = np.zeros((1, num_cols))
Ir = np.identity(num_rows)
Ic = np.identity(num_cols)
I1 = np.identity(1)
assert_equal(matrix_normal.rvs(mean=M, rowcov=U, colcov=V).shape,
(num_rows, num_cols))
assert_equal(matrix_normal.rvs(mean=M).shape,
(num_rows, num_cols))
assert_equal(matrix_normal.rvs(rowcov=U).shape,
(num_rows, 1))
assert_equal(matrix_normal.rvs(colcov=V).shape,
(1, num_cols))
assert_equal(matrix_normal.rvs(mean=M, colcov=V).shape,
(num_rows, num_cols))
assert_equal(matrix_normal.rvs(mean=M, rowcov=U).shape,
(num_rows, num_cols))
assert_equal(matrix_normal.rvs(rowcov=U, colcov=V).shape,
(num_rows, num_cols))
assert_equal(matrix_normal(mean=M).rowcov, Ir)
assert_equal(matrix_normal(mean=M).colcov, Ic)
assert_equal(matrix_normal(rowcov=U).mean, Zr)
assert_equal(matrix_normal(rowcov=U).colcov, I1)
assert_equal(matrix_normal(colcov=V).mean, Zc)
assert_equal(matrix_normal(colcov=V).rowcov, I1)
assert_equal(matrix_normal(mean=M, rowcov=U).colcov, Ic)
assert_equal(matrix_normal(mean=M, colcov=V).rowcov, Ir)
assert_equal(matrix_normal(rowcov=U, colcov=V).mean, Z)
def test_moments(self):
# Check that the sample moments match the parameters
num_rows = 4
num_cols = 3
M = 0.3 * np.ones((num_rows,num_cols))
U = 0.5 * np.identity(num_rows) + 0.5 * np.ones((num_rows, num_rows))
V = 0.7 * np.identity(num_cols) + 0.3 * np.ones((num_cols, num_cols))
N = 1000
frozen = matrix_normal(mean=M, rowcov=U, colcov=V)
X = frozen.rvs(size=N, random_state=1234)
sample_mean = np.mean(X,axis=0)
assert_allclose(sample_mean, M, atol=0.1)
sample_colcov = np.cov(X.reshape(N*num_rows,num_cols).T)
assert_allclose(sample_colcov, V, atol=0.1)
sample_rowcov = np.cov(np.swapaxes(X,1,2).reshape(
N*num_cols,num_rows).T)
assert_allclose(sample_rowcov, U, atol=0.1)
def test_moments(self):
# Check that the sample moments match the parameters
num_rows = 4
num_cols = 3
M = 0.3 * np.ones((num_rows, num_cols))
U = 0.5 * np.identity(num_rows) + 0.5 * np.ones((num_rows, num_rows))
V = 0.7 * np.identity(num_cols) + 0.3 * np.ones((num_cols, num_cols))
N = 1000
frozen = matrix_normal(mean=M, rowcov=U, colcov=V)
X = frozen.rvs(size=N, random_state=1234)
sample_mean = np.mean(X, axis=0)
assert_allclose(sample_mean, M, atol=0.1)
sample_colcov = np.cov(X.reshape(N * num_rows, num_cols).T)
assert_allclose(sample_colcov, V, atol=0.1)
sample_rowcov = np.cov(
np.swapaxes(X, 1, 2).reshape(N * num_cols, num_rows).T)
assert_allclose(sample_rowcov, U, atol=0.1)
def test_matches_multivariate(self):
# Check that the pdfs match those obtained by vectorising and
# treating as a multivariate normal.
for i in range(1, 5):
for j in range(1, 5):
M = 0.3 * np.ones((i, j))
U = 0.5 * np.identity(i) + 0.5 * np.ones((i, i))
V = 0.7 * np.identity(j) + 0.3 * np.ones((j, j))
frozen = matrix_normal(mean=M, rowcov=U, colcov=V)
X = frozen.rvs(random_state=1234)
pdf1 = frozen.pdf(X)
logpdf1 = frozen.logpdf(X)
vecX = X.T.flatten()
vecM = M.T.flatten()
cov = np.kron(V, U)
pdf2 = multivariate_normal.pdf(vecX, mean=vecM, cov=cov)
logpdf2 = multivariate_normal.logpdf(vecX, mean=vecM, cov=cov)
assert_allclose(pdf1, pdf2, rtol=1E-10)
assert_allclose(logpdf1, logpdf2, rtol=1E-10)
def test_array_input(self):
# Check array of inputs has the same output as the separate entries.
num_rows = 4
num_cols = 3
M = 0.3 * np.ones((num_rows,num_cols))
U = 0.5 * np.identity(num_rows) + 0.5 * np.ones((num_rows, num_rows))
V = 0.7 * np.identity(num_cols) + 0.3 * np.ones((num_cols, num_cols))
N = 10
frozen = matrix_normal(mean=M, rowcov=U, colcov=V)
X1 = frozen.rvs(size=N, random_state=1234)
X2 = frozen.rvs(size=N, random_state=4321)
X = np.concatenate((X1[np.newaxis,:,:,:],X2[np.newaxis,:,:,:]), axis=0)
assert_equal(X.shape, (2, N, num_rows, num_cols))
array_logpdf = frozen.logpdf(X)
assert_equal(array_logpdf.shape, (2, N))
for i in range(2):
for j in range(N):
separate_logpdf = matrix_normal.logpdf(X[i,j], mean=M,
rowcov=U, colcov=V)
assert_allclose(separate_logpdf, array_logpdf[i,j], 1E-10)
def test_matches_multivariate(self):
# Check that the pdfs match those obtained by vectorising and
# treating as a multivariate normal.
for i in range(1,5):
for j in range(1,5):
M = 0.3 * np.ones((i,j))
U = 0.5 * np.identity(i) + 0.5 * np.ones((i,i))
V = 0.7 * np.identity(j) + 0.3 * np.ones((j,j))
frozen = matrix_normal(mean=M, rowcov=U, colcov=V)
X = frozen.rvs(random_state=1234)
pdf1 = frozen.pdf(X)
logpdf1 = frozen.logpdf(X)
vecX = X.T.flatten()
vecM = M.T.flatten()
cov = np.kron(V,U)
pdf2 = multivariate_normal.pdf(vecX, mean=vecM, cov=cov)
logpdf2 = multivariate_normal.logpdf(vecX, mean=vecM, cov=cov)
assert_allclose(pdf1, pdf2, rtol=1E-10)
assert_allclose(logpdf1, logpdf2, rtol=1E-10)
def test_frozen_matrix_normal(self):
for i in range(1,5):
for j in range(1,5):
M = 0.3 * np.ones((i,j))
U = 0.5 * np.identity(i) + 0.5 * np.ones((i,i))
V = 0.7 * np.identity(j) + 0.3 * np.ones((j,j))
frozen = matrix_normal(mean=M, rowcov=U, colcov=V)
rvs1 = frozen.rvs(random_state=1234)
rvs2 = matrix_normal.rvs(mean=M, rowcov=U, colcov=V,
random_state=1234)
assert_equal(rvs1, rvs2)
X = frozen.rvs(random_state=1234)
pdf1 = frozen.pdf(X)
pdf2 = matrix_normal.pdf(X, mean=M, rowcov=U, colcov=V)
assert_equal(pdf1, pdf2)
logpdf1 = frozen.logpdf(X)
logpdf2 = matrix_normal.logpdf(X, mean=M, rowcov=U, colcov=V)
assert_equal(logpdf1, logpdf2)
def update_slds_theta(Y, fwd_vals, params, priors, ar=1, **kwargs):
S_0, n_0 = priors
L, D, theta = params['L'], params['D'], params['theta']
n = fwd_vals['n']
# set pseudo_obs = pseudo_obs
if ar == 2:
if D > 1:
mx, mn = np.max(Y.shape), np.min(Y.shape)
Y_bar = np.concatenate(
[np.zeros_like(Y), np.zeros_like(Y)],
axis=-1).astype(np.float32)
Y_bar[0, :] += np.concatenate([Y[0, :], Y[0, :]])
Y_bar[1:, :mn] += Y[:-1]
Y_bar[1, mn:] += Y[0, :]
Y_bar[2:, mn:] += Y[:-2]
else:
Y_bar = np.zeros(shape=(Y.shape[0], D * 2), dtype=np.float32)
Y_bar[0] += Y[0]
Y_bar[1:, 0] += Y[:-1]
Y_bar[1, 1] += Y[0]
Y_bar[2:, 1] += Y[:-2]
else:
# lagged observations
Y_bar = np.zeros_like(Y, dtype=np.float32)
Y_bar[0] += Y[0]
Y_bar[1:] += Y[:-1]
# step 5 caluculate the sufficient statistics for the pseudo_obs
S_ybarybar, S_yybar, S_yy = slds_sufficient_statistics(
Y, Y_bar, fwd_vals, ar, params)
for k in range(0, L):
S_ybarybar_k, S_yybar_k, S_yy_k = S_ybarybar[k], S_yybar[k], S_yy[k]
S_ybarybar_k_inv = invert(S_ybarybar_k)
# sigma_k
Sy_vert_ybar = S_yy_k - np.dot(np.dot(S_yybar_k, S_ybarybar_k_inv),
S_yybar_k.T)
# sig = stats.invwishart(scale=Sy_vert_ybar+S_0, df=np.sum(n[:,k])+n_0).rvs().astype(np.float32)
sig = stats.invwishart(scale=Sy_vert_ybar + S_0,
df=np.sum(n[k]) + n_0).rvs().astype(np.float32)
if type(sig) != np.ndarray:
sig = np.reshape(sig, newshape=(1, 1))
sigma_ = theta[k]['sigma']
# sigma_ = sig
sigma_inv_ = invert(sigma_)
# TODO: I there is ambiguity in paper, not sure about this
M = np.dot(S_yybar_k, S_ybarybar_k_inv) # PhD suggests this one
# K_inv = S_ybarybar_k_inv
if ar == 2:
A = stats.matrix_normal(mean=M,
rowcov=sigma_inv_,
colcov=S_ybarybar_k_inv).rvs()
A = np.array([A[:D, D:], A[:D, :D]], dtype=np.float32)
else:
A = stats.matrix_normal(mean=M,
rowcov=sigma_inv_,
colcov=S_ybarybar_k_inv).rvs().reshape(
D, -D).astype(np.float32)
theta[k]['A'] = A
theta[k]['sigma'] = sig
return theta