class MLSVGP(BASESVGP):
def __init__(self,
dim_in,
dim_out,
kern,
likelihood,
dim_h,
num_h,
feat=None,
mean_function=None,
q_diag=False,
whiten=True,
Z=None,
num_data=None,
observed_config_space_dim=0,
latent_to_conf_space_kernel=None,
latent_to_conf_space_likelihood=None,
**kwargs):
super(MLSVGP, self).__init__(dim_in=dim_in,
dim_out=dim_out,
kern=kern,
likelihood=likelihood,
feat=feat,
mean_function=mean_function,
q_diag=q_diag,
whiten=whiten,
Z=Z,
num_data=num_data,
**kwargs)
self.dim_h = dim_h
self.num_h = num_h
self.observed_config_space = observed_config_space_dim
self.mean_psi = Zero(observed_config_space_dim)
self.configuration_kernel = latent_to_conf_space_kernel
self.configuration_likelihood = latent_to_conf_space_likelihood
# Initialize task variables
H_mu = np.random.randn(num_h, dim_h)
H_var = np.log(np.ones_like(H_mu) * 0.1)
H_init = np.hstack([H_mu, H_var])
self.H = Parameter(H_init, dtype=settings.float_type, name="H")
# Create placeholders
self.H_ids_ph = tf.placeholder(tf.int32, [None])
self.H_unique_ph = tf.placeholder(tf.int32, [None])
self.H_scale = tf.placeholder(settings.float_type, [])
self.psi_ph = tf.placeholder(dtype=settings.float_type,
shape=[None, self.observed_config_space])
@params_as_tensors
def build_likelihood(self):
# Get prior KL.
KL_U = self.build_prior_KL()
H_sample, KL_H = self.sample_qH(self.H)
KL_H = tf.reduce_sum(tf.gather(KL_H, self.H_unique_ph))
KL_H *= self.H_scale
H_sample_Yph = tf.gather(H_sample, self.H_ids_ph)
H_sample_psi = tf.gather(H_sample, self.H_unique_ph)
XH = tf.concat([self.X_mu_ph, H_sample_Yph], 1)
# Get conditionals
fmean, fvar = self._build_predict(XH, full_cov=False)
# Get variational expectations.
var_exp_Y = self.likelihood.variational_expectations(
fmean, fvar, self.Y_ph)
lik_term = (tf.reduce_sum(var_exp_Y) +
self._build_likelihood_psi(H_sample_psi)) * self.data_scale
likelihood = lik_term - KL_U - KL_H
return likelihood
@params_as_tensors
def _build_predict(self, XHnew, full_cov=False, full_output_cov=False):
mu, var = conditional(XHnew,
self.feature,
self.kern,
self.q_mu,
q_sqrt=self.q_sqrt,
full_cov=full_cov,
white=self.whiten,
full_output_cov=full_output_cov)
return mu + self.mean_function(XHnew), var
@params_as_tensors
def _build_likelihood_psi(self, H_sample):
r"""
Construct configuration tensorflow function to compute the likelihood.
\log p(Y | theta).
"""
K = self.configuration_kernel.K(H_sample) + tf.eye(
tf.shape(H_sample)[0],
dtype=settings.float_type) * self.configuration_likelihood.variance
L = tf.cholesky(K)
m = self.mean_psi(H_sample)
logpdf = multivariate_normal(
self.psi_ph, m,
L) # (R,) log-likelihoods for each independent dimension of Y
return tf.reduce_sum(logpdf)
@params_as_tensors
def build_predict_psi(self, Hnew, full_cov=False):
H_sample = tf.gather(self.H[:, :self.dim_h], self.H_unique_ph)
y = self.psi_ph - self.mean_psi(H_sample)
Kmn = self.configuration_kernel.K(H_sample, Hnew)
Kmm_sigma = self.configuration_kernel.K(H_sample) + tf.eye(
tf.shape(H_sample)[0],
dtype=settings.float_type) * self.configuration_likelihood.variance
Knn = self.configuration_kernel.K(
Hnew) if full_cov else self.configuration_kernel.Kdiag(Hnew)
f_mean, f_var = base_conditional(
Kmn, Kmm_sigma, Knn, y, full_cov=full_cov,
white=False) # N x P, N x P or P x N x N
return f_mean + self.mean_psi(Hnew), f_var
@params_as_tensors
def _build_predict_uncertain(self,
Xnew_mu,
Xnew_var,
full_cov=False,
full_output_cov=False,
Luu=None):
H = tf.gather(self.H, self.H_ids_ph)
H_mu = H[:, :self.dim_h]
H_var = tf.matrix_diag(tf.exp(H[:, self.dim_h:]))
XH_mu = tf.concat([Xnew_mu, H_mu], 1)
XH_var = block_diag(Xnew_var[0], H_var[0])[None, :, :]
mu, var, inp_out_cov = uncertain_conditional(
Xnew_mu=XH_mu,
Xnew_var=XH_var,
feat=self.feature,
kern=self.kern,
q_mu=self.q_mu,
q_sqrt=self.q_sqrt,
Luu=Luu,
mean_function=self.mean_function,
full_cov=full_cov,
full_output_cov=full_output_cov,
white=self.whiten)
return mu, var, inp_out_cov[:, :-self.dim_h]
@params_as_tensors
def build_predict_uncertain(self,
XH_mu,
XH_var,
full_cov=False,
full_output_cov=False,
Luu=None):
mu, var, inp_out_cov = uncertain_conditional(
Xnew_mu=XH_mu,
Xnew_var=XH_var,
feat=self.feature,
kern=self.kern,
q_mu=self.q_mu,
q_sqrt=self.q_sqrt,
Luu=Luu,
mean_function=self.mean_function,
full_cov=full_cov,
full_output_cov=full_output_cov,
white=self.whiten)
return mu, var, inp_out_cov[:, :-self.dim_h]
@params_as_tensors
def sample_qH(self, H):
h_mu = H[:, :self.dim_h]
h_var = tf.exp(H[:, self.dim_h:])
qh = dist.Normal(h_mu, tf.sqrt(h_var))
ph = dist.Normal(tf.zeros_like(h_mu), tf.ones_like(h_var))
kl_h = dist.kl_divergence(qh, ph)
h_sample = qh.sample()
return h_sample, kl_h
@params_as_tensors
def compute_Luu(self):
Kuu = self.feature.Kuu(self.kern,
jitter=settings.numerics.jitter_level)
return tf.cholesky(Kuu)
def get_H_space(self, session):
H = self.H.read_value(session=session)
H_mu = H[:, :self.dim_h]
H_var = np.exp(H[:, self.dim_h:])
return H_mu, H_var
def get_model_param(self, session):
lik_noise = self.likelihood.variance.read_value(session=session)
kern_var = self.kern.variance.read_value(session=session)
kern_ls = self.kern.lengthscales.read_value(session=session)
return lik_noise, kern_var, kern_ls
def get_H_subset(self, session, end_task_id, start_task_id=0):
H = self.H.read_value(session=session)[start_task_id:end_task_id]
H[:, self.dim_h:] = np.exp(H[:, self.dim_h:])
return H
class MLSVGP(BASESVGP):
def __init__(self,
dim_in, dim_out,
kern, likelihood,
dim_h, num_h,
feat=None,
mean_function=None,
q_diag=False,
whiten=True,
Z=None,
num_data=None,
**kwargs):
super(MLSVGP, self).__init__(
dim_in=dim_in, dim_out=dim_out,
kern=kern, likelihood=likelihood,
feat=feat,
mean_function=mean_function,
q_diag=q_diag,
whiten=whiten,
Z=Z,
num_data=num_data,
**kwargs)
self.dim_h = dim_h
self.num_h = num_h
# Initialize task variables
H_mu = np.random.randn(num_h, dim_h)
H_var = np.log(np.ones_like(H_mu) * 0.1)
H_init = np.hstack([H_mu, H_var])
self.H = Parameter(H_init, dtype=settings.float_type, name="H")
# Create placeholders
self.H_ids_ph = tf.placeholder(tf.int32, [None])
self.H_unique_ph = tf.placeholder(tf.int32, [None])
self.H_scale = tf.placeholder(settings.float_type, [])
@params_as_tensors
def _build_likelihood(self):
# Get prior KL.
KL_U = self.build_prior_KL()
H_sample, KL_H = self.sample_qH(self.H)
KL_H = tf.reduce_sum(tf.gather(KL_H, self.H_unique_ph))
KL_H *= self.H_scale
H_sample = tf.gather(H_sample, self.H_ids_ph)
XH = tf.concat([self.X_mu_ph, H_sample], 1)
# Get conditionals
fmean, fvar = self._build_predict(XH, full_cov=False)
# Get variational expectations.
var_exp = self.likelihood.variational_expectations(fmean, fvar, self.Y_ph)
lik_term = tf.reduce_sum(var_exp) * self.data_scale
likelihood = lik_term - KL_U - KL_H
return likelihood
@params_as_tensors
def _build_predict(self, XHnew, full_cov=False, full_output_cov=False):
mu, var = conditional(XHnew, self.feature, self.kern, self.q_mu, q_sqrt=self.q_sqrt, full_cov=full_cov,
white=self.whiten, full_output_cov=full_output_cov)
return mu + self.mean_function(XHnew), var
@params_as_tensors
def _build_predict_uncertain(self, Xnew_mu, Xnew_var,
full_cov=False, full_output_cov=False, Luu=None):
H = tf.gather(self.H, self.H_ids_ph)
H_mu = H[:, :self.dim_h]
H_var = tf.matrix_diag(tf.exp(H[:, self.dim_h:]))
XH_mu = tf.concat([Xnew_mu, H_mu], 1)
XH_var = block_diag(Xnew_var[0], H_var[0])[None, :, :]
mu, var, inp_out_cov = uncertain_conditional(
Xnew_mu=XH_mu, Xnew_var=XH_var, feat=self.feature, kern=self.kern,
q_mu=self.q_mu, q_sqrt=self.q_sqrt, Luu=Luu,
mean_function=self.mean_function,
full_cov=full_cov, full_output_cov=full_output_cov, white=self.whiten)
return mu, var, inp_out_cov[:, :-self.dim_h]
@params_as_tensors
def sample_qH(self, H):
h_mu = H[:, :self.dim_h]
h_var = tf.exp(H[:, self.dim_h:])
qh = dist.Normal(h_mu, tf.sqrt(h_var))
ph = dist.Normal(tf.zeros_like(h_mu), tf.ones_like(h_var))
kl_h = dist.kl_divergence(qh, ph)
h_sample = qh.sample()
return h_sample, kl_h
@params_as_tensors
def compute_Luu(self):
Kuu = self.feature.Kuu(self.kern, jitter=settings.numerics.jitter_level)
return tf.cholesky(Kuu)
def get_H_space(self, session):
H = self.H.read_value(session=session)
H_mu = H[:, :self.dim_h]
H_var = np.exp(H[:, self.dim_h:])
return H_mu, H_var
def get_model_param(self, session):
lik_noise = self.likelihood.variance.read_value(session=session)
kern_var = self.kern.variance.read_value(session=session)
kern_ls = self.kern.lengthscales.read_value(session=session)
return lik_noise, kern_var, kern_ls