def proposal(self, current, current_log_pdf, **kwargs):
if self.Z is None and self.schedule is None:
raise ValueError("%s has not seen data yet. Call set_batch()" %
self.__class__.__name__)
if current_log_pdf is None:
current_log_pdf = self.target_log_pdf(current)
L_R = self._construct_proposal_covariance(current)
proposal = sample_gaussian(N=1,
mu=current,
Sigma=L_R,
is_cholesky=True)[0]
proposal_log_prob = log_gaussian_pdf(proposal,
current,
L_R,
is_cholesky=True)
proposal_log_pdf = self.target_log_pdf(proposal)
# probability of proposing y when would be sitting at proposal
L_R_inv = self._construct_proposal_covariance(proposal)
proopsal_log_prob_inv = log_gaussian_pdf(current,
proposal,
L_R_inv,
is_cholesky=True)
log_acc_prob = proposal_log_pdf - current_log_pdf + proopsal_log_prob_inv - proposal_log_prob
log_acc_prob = np.min([0, log_acc_prob])
results_kwargs = {}
return proposal, proposal_log_pdf, current_log_pdf, proposal_log_prob, proopsal_log_prob_inv, results_kwargs
def proposal(self, current):
"""
Returns a sample from the proposal centred at current, and its log-probability
"""
if self.schedule is None and (self.mu is None or self.C is None):
raise ValueError("Kameleon has not seen data yet." \
"Either call set_batch_covariance() or set update schedule")
L_R = self._construct_proposal_covariance(current)
proposal = sample_gaussian(N=1,
mu=current,
Sigma=L_R,
is_cholesky=True)[0]
proposal_log_prob = log_gaussian_pdf(proposal,
current,
L_R,
is_cholesky=True)
# probability of proposing current when would be sitting at proposal
L_R_inv = self._construct_proposal_covariance(proposal)
proopsal_log_prob_inv = log_gaussian_pdf(current,
proposal,
L_R_inv,
is_cholesky=True)
return proposal, proposal_log_prob, proopsal_log_prob_inv
def test_log_gaussian_pdf_multiple_equals_log_gaussian_pdf_looped_isotropic_cov(
):
N = 100
D = 3
X = np.random.randn(N, D)
mu = np.random.randn(D)
cov_scaling = 2.
log_pdfs = log_gaussian_pdf_multiple(X, mu, cov_scaling=cov_scaling)
grads = log_gaussian_pdf_multiple(X,
mu,
compute_grad=True,
cov_scaling=cov_scaling)
print log_pdfs
for i, x in enumerate(X):
log_pdf = log_gaussian_pdf(x, mu, cov_scaling=cov_scaling)
grad = log_gaussian_pdf(x,
mu,
compute_grad=True,
cov_scaling=cov_scaling)
assert_allclose(log_pdf, log_pdfs[i])
assert_allclose(grad, grads[i])
def proposal(self, current):
"""
Returns a sample from the proposal centred at current, and its log-probability
"""
if self.schedule is None and (self.mu is None or self.C is None):
raise ValueError("Kameleon has not seen data yet." \
"Either call set_batch_covariance() or set update schedule")
L_R = self._construct_proposal_covariance(current)
proposal = sample_gaussian(N=1, mu=current, Sigma=L_R, is_cholesky=True)[0]
proposal_log_prob = log_gaussian_pdf(proposal, current, L_R, is_cholesky=True)
# probability of proposing current when would be sitting at proposal
L_R_inv = self._construct_proposal_covariance(proposal)
proopsal_log_prob_inv = log_gaussian_pdf(current, proposal, L_R_inv, is_cholesky=True)
return proposal, proposal_log_prob, proopsal_log_prob_inv
def test_log_gaussian_pdf_multiple_equals_log_gaussian_pdf_looped_full_cov():
N = 100
D = 3
X = np.random.randn(N, D)
mu = np.random.randn(D)
L_C = np.linalg.cholesky(np.dot(X.T, X) + np.eye(D))
cov_scaling = 2.
log_pdfs = log_gaussian_pdf_multiple(X, mu, L_C, is_cholesky=True, cov_scaling=cov_scaling)
grads = log_gaussian_pdf_multiple(X, mu, L_C, is_cholesky=True, compute_grad=True, cov_scaling=cov_scaling)
print log_pdfs
for i, x in enumerate(X):
log_pdf = log_gaussian_pdf(x, mu, L_C, is_cholesky=True, cov_scaling=cov_scaling)
grad = log_gaussian_pdf(x, mu, L_C, is_cholesky=True, compute_grad=True, cov_scaling=cov_scaling)
assert_allclose(log_pdf, log_pdfs[i])
assert_allclose(grad, grads[i])
def test_log_gaussian_pdf_multiple_equals_log_gaussian_pdf_looped_isotropic_cov():
N = 100
D = 3
X = np.random.randn(N, D)
mu = np.random.randn(D)
cov_scaling = 2.
log_pdfs = log_gaussian_pdf_multiple(X, mu, cov_scaling=cov_scaling)
grads = log_gaussian_pdf_multiple(X, mu, compute_grad=True, cov_scaling=cov_scaling)
print log_pdfs
for i, x in enumerate(X):
log_pdf = log_gaussian_pdf(x, mu, cov_scaling=cov_scaling)
grad = log_gaussian_pdf(x, mu, compute_grad=True, cov_scaling=cov_scaling)
assert_allclose(log_pdf, log_pdfs[i])
assert_allclose(grad, grads[i])
def proposal(self, current, current_log_pdf, **kwargs):
if self.Z is None and self.schedule is None:
raise ValueError("%s has not seen data yet. Call set_batch()" % self.__class__.__name__)
if current_log_pdf is None:
current_log_pdf = self.target_log_pdf(current)
L_R = self._construct_proposal_covariance(current)
proposal = sample_gaussian(N=1, mu=current, Sigma=L_R, is_cholesky=True)[0]
proposal_log_prob = log_gaussian_pdf(proposal, current, L_R, is_cholesky=True)
proposal_log_pdf = self.target_log_pdf(proposal)
# probability of proposing y when would be sitting at proposal
L_R_inv = self._construct_proposal_covariance(proposal)
proopsal_log_prob_inv = log_gaussian_pdf(current, proposal, L_R_inv, is_cholesky=True)
log_acc_prob = proposal_log_pdf - current_log_pdf + proopsal_log_prob_inv - proposal_log_prob
log_acc_prob = np.min([0, log_acc_prob])
results_kwargs = {}
return proposal, proposal_log_pdf, current_log_pdf, proposal_log_prob, proopsal_log_prob_inv, results_kwargs
def test_log_gaussian_pdf_multiple_equals_log_gaussian_pdf_looped_full_cov():
N = 100
D = 3
X = np.random.randn(N, D)
mu = np.random.randn(D)
L_C = np.linalg.cholesky(np.dot(X.T, X) + np.eye(D))
cov_scaling = 2.
log_pdfs = log_gaussian_pdf_multiple(X,
mu,
L_C,
is_cholesky=True,
cov_scaling=cov_scaling)
grads = log_gaussian_pdf_multiple(X,
mu,
L_C,
is_cholesky=True,
compute_grad=True,
cov_scaling=cov_scaling)
print log_pdfs
for i, x in enumerate(X):
log_pdf = log_gaussian_pdf(x,
mu,
L_C,
is_cholesky=True,
cov_scaling=cov_scaling)
grad = log_gaussian_pdf(x,
mu,
L_C,
is_cholesky=True,
compute_grad=True,
cov_scaling=cov_scaling)
assert_allclose(log_pdf, log_pdfs[i])
assert_allclose(grad, grads[i])