def buildZ(self, normalize=False):
Z = []
self.k_gamma = 0
for ran in self.ran_list:
id, dims = ran
values = []
if id is None:
values = np.ones(self.N)
else:
values = rutils.repeat(self.covariates[id][0],
self.covariates[id][1], self.dimensions)
self.k_gamma += np.prod(dims[self.n_grouping_dims:])
Z.append(
values.reshape((-1, 1)) * np.tile(
rutils.kronecker(dims[self.n_grouping_dims:],
self.dimensions, self.n_grouping_dims),
(self.n_groups, 1)))
if self.k_gamma > 0:
self.Z = np.hstack(Z)
col_norm = np.linalg.norm(self.Z, axis=0)
if normalize:
print('normalizing Z ...')
print(col_norm)
self.Z = self.Z / col_norm
return col_norm
else:
self.Z = np.zeros((self.N, 1))
return 0.0
def test_draw_random_only(self, random_effects):
np.random.seed(127)
dimensions = [9, 3, 2, 2]
N = np.prod(dimensions)
Y_true = np.zeros(N)
dct = {}
for i, effect in enumerate(random_effects):
Z = rutils.kronecker(effect, dimensions, 0)
u = np.random.randn(np.prod(effect))
Y_true += Z.dot(u)
dct['intercept' + str(i)] = ([
effect[j] == dimensions[j] for j in range(len(dimensions))
], None)
delta_true = .005
Y = Y_true + np.random.randn(N) * np.sqrt(delta_true)
model = LME(dimensions, 1, Y, {}, {}, {}, False, random_effects=dct)
model.optimize(inner_print_level=0)
model.postVarRandom()
n_draws = 1000
_, u_samples = model.draw(n_draws=n_draws)
u1 = np.concatenate(
[u[:np.prod(random_effects[0][1:])] for u in model.u_soln])
u1_sample_mean = np.mean(u_samples[0].reshape((-1, n_draws)), axis=1)
assert np.linalg.norm(u1 - u1_sample_mean) / np.linalg.norm(u1) < .05
u2 = np.concatenate([
u[np.prod(random_effects[0][1:]):np.prod(random_effects[0][1:]) +
np.prod(random_effects[1][1:])] for u in model.u_soln
])
u2_sample_mean = np.mean(u_samples[1].reshape((-1, n_draws)), axis=1)
assert np.linalg.norm(u2 - u2_sample_mean) / np.linalg.norm(u2) < .05
model.outputDraws()
return
def test_random_effect_with_gaussian_prior(self, random_effect, sd):
np.random.seed(127)
dimensions = [200, 2, 3, 2]
N = np.prod(dimensions)
Y_true = np.zeros(N)
Z = rutils.kronecker(random_effect, dimensions, 0)
u = np.random.randn(np.prod(random_effect)) * .5
dct1 = {
'intercept': ([
random_effect[j] == dimensions[j]
for j in range(len(dimensions))
], None)
}
dct2 = {
'intercept': ([
random_effect[j] == dimensions[j]
for j in range(len(dimensions))
], sd)
}
delta_true = 0.005
Y_true += Z.dot(u)
Y = Y_true + np.random.randn(N) * np.sqrt(delta_true)
model1 = LME(dimensions, 1, Y, {}, {}, {}, False, random_effects=dct1)
model1.optimize(inner_print_level=0)
gamma1 = model1.gamma_soln
u_var1 = np.var(model1.u_soln)
model2 = LME(dimensions, 1, Y, {}, {}, {}, False, random_effects=dct2)
model2.optimize(inner_print_level=0)
gamma2 = model2.gamma_soln
u_var2 = np.var(model2.u_soln)
assert all(gamma1 > gamma2)
assert u_var1 > u_var2
def test_repeat_kron(self, full_dims, dims):
values = np.random.randn(np.prod(dims))
Z = rutils.kronecker(dims, full_dims, 0)
vals = rutils.repeat(values, dims, full_dims)
y = np.random.randn(np.prod(full_dims))
vals2 = rutils.repeatTranspose(y, dims, full_dims)
assert (np.linalg.norm(Z.dot(values) - vals) < 1e-10
and np.linalg.norm(np.transpose(Z).dot(y) - vals2) < 1e-10)
def test_draw(self, dimensions, random_effects):
np.random.seed(127)
#dimensions = [9, 3, 2, 2]
N = np.prod(dimensions)
X = np.ones((N, 2))
X[:, 1] = np.random.randn(N)
beta_true = [1., -0.6]
Y_true = X.dot(beta_true)
dct = {}
for i, effect in enumerate(random_effects):
Z = rutils.kronecker(effect, dimensions, 0)
u = np.random.randn(np.prod(effect))
Y_true += Z.dot(u)
dct['intercept' + str(i)] = ([
effect[j] == dimensions[j] for j in range(len(dimensions))
], None)
delta_true = .005
Y = Y_true + np.random.randn(N) * np.sqrt(delta_true)
model = LME(dimensions,
1,
Y, {'cov': (X[:, 1], [True] * len(dimensions))}, {},
{'cov': [-float('inf'), float('inf')]},
True,
random_effects=dct)
model.optimize(inner_print_level=0)
model.postVarGlobal()
if len(random_effects) > 0:
model.postVarRandom()
n_draws = 1000
beta_samples, u_samples = model.draw(n_draws=n_draws)
beta_sample_mean = np.mean(beta_samples, axis=1)
assert np.linalg.norm(beta_sample_mean - model.beta_soln
) / np.linalg.norm(model.beta_soln) < .02
if len(random_effects) > 0:
u1 = np.concatenate(
[u[:np.prod(random_effects[0][1:])] for u in model.u_soln])
u1_sample_mean = np.mean(u_samples[0].reshape((-1, n_draws)),
axis=1)
assert np.linalg.norm(u1 -
u1_sample_mean) / np.linalg.norm(u1) < .05
u2 = np.concatenate([
u[np.prod(random_effects[0][1:]
):np.prod(random_effects[0][1:]) +
np.prod(random_effects[1][1:])] for u in model.u_soln
])
u2_sample_mean = np.mean(u_samples[1].reshape((-1, n_draws)),
axis=1)
assert np.linalg.norm(u2 -
u2_sample_mean) / np.linalg.norm(u2) < .05
model.outputDraws()
def _postVarGlobal(self):
"""
y_k = X_k beta + Z_k u_k + epsilon, u has var_mat D, epsilon has var_mat R
Var(beta) = inv( sum_k X_k' inv(Z_k*D*Z_k' + R) X_k )
"""
assert self.k_beta > 0
Z_split = np.split(self.Z, self.n_groups)
Z_split = np.split(np.zeros((self.N, self.k_gamma)), self.n_groups)
self.var_beta = np.zeros((self.k_beta, self.k_beta))
X = np.zeros((self.N, self.k_beta))
start = 0
if self.global_intercept is True:
X[:, start] = np.ones(self.N)
start += 1
for i in range(len(self.global_ids)):
ind = self.global_ids[i]
values, dims = self.covariates[ind]
assert values.shape[0] == np.prod(dims)
X[:, start] = rutils.repeat(values, dims, self.dimensions)
start += 1
for indicator in self.indicators:
X[:, start:start + np.prod(indicator)] = rutils.kronecker(
indicator, self.dimensions, 0)
start += np.prod(indicator)
X_split = np.split(X, self.n_groups)
S2 = []
if self.S is None:
if self.share_obs_std is True:
S2 = np.ones(self.N) * self.delta_soln
else:
S2 = np.repeat(self.delta_soln, self.grouping)
else:
S2 = self.S**2
S2_split = np.split(S2, self.n_groups)
for i in range(self.n_groups):
V = Z_split[i].dot(np.diag(self.gamma_soln)).dot(np.transpose(Z_split[i])) \
+ S2_split[i]*np.identity(self.grouping[i])
self.var_beta += np.transpose(X_split[i]).dot(
np.linalg.inv(V)).dot(X_split[i])
self.var_beta = np.linalg.inv(self.var_beta)
def test_indicators(self, dimensions, indicator):
"""
Test if indicator matrix is built correctly.
"""
dct = {
'intercept':
[indicator[j] == dimensions[j] for j in range(len(dimensions))]
}
y = np.random.randn(np.prod(dimensions))
model = LME(dimensions, 0, y, {}, dct, {}, False, {})
Z = rutils.kronecker(indicator, dimensions, 0)
x = np.random.randn(np.prod(indicator))
assert (np.linalg.norm(model.X(x) - Z.dot(x)) < 1e-10) and \
(np.linalg.norm(model.XT(y) - np.transpose(Z).dot(y)) < 1e-10)
def test_random_intercept(self, dimensions, random_intercept):
"""
Test if random intercept matrix is built correctly.
"""
dct = {
'intercept': ([
random_intercept[j] == dimensions[j]
for j in range(len(dimensions))
], None)
}
y = np.random.randn(np.prod(dimensions))
model = LME(dimensions, 1, y, {}, {}, {}, True, dct)
Z = np.tile(rutils.kronecker(random_intercept[1:], dimensions, 1),
(dimensions[0], 1))
model.buildZ()
assert np.linalg.norm(Z - model.Z) == 0.0
def test_repeat_covariate(self, dimensions, cov_dim):
N = np.prod(dimensions)
X = np.ones((N, 2)) # 1st column is intercept
cov = np.random.randn(np.prod(cov_dim))
cov_dim_bool = [
cov_dim[i] == dimensions[i] for i in range(len(dimensions))
]
Z = rutils.kronecker(cov_dim, dimensions, 0)
X[:, 1] = Z.dot(cov)
beta_true = [1., -0.6] # beta_0 for intercept
Y = X.dot(beta_true)
model = LME(dimensions, 0, Y, {'cov1': (cov, cov_dim_bool)}, {},
{'cov1': [-float('inf'), float('inf')]}, True, {})
beta = np.random.randn(2)
assert np.linalg.norm(model.X(beta) - X.dot(beta)) < 1e-10
y = np.random.randn(N)
assert np.linalg.norm(model.XT(y) - np.transpose(X).dot(y)) < 1e-10
model._buildX()
assert np.linalg.norm(model.Xm - X) < 1e-10
def _buildX(self):
X = np.zeros((self.N, self.k_beta))
start = 0
if self.global_intercept is True:
X[:, start] = np.ones(self.N)
start += 1
for i in range(len(self.global_ids)):
ind = self.global_ids[i]
values, dims = self.covariates[ind]
assert values.shape[0] == np.prod(dims)
X[:, start] = rutils.repeat(values, dims, self.dimensions)
start += 1
for indicator in self.indicators:
X[:, start:start + np.prod(indicator)] = rutils.kronecker(
indicator, self.dimensions, 0)
start += np.prod(indicator)
self.Xm = X
