def logprob(self, x, model):
"""compute the log probability of observations x
This includes the model likelihood as well as any prior
probability of the parameters
Returns
-------
lp : float
the log probability
"""
# set values of the parameers in self.params to be x
hyperparameter_utils.set_params_from_array(self.params, x)
lp = 0.0
# sum the log probabilities of the parameter priors
for param in self.params:
lp += param.prior_logprob()
if np.isnan(lp): # Positive infinity should be ok, right?
print('Param diagnostics:')
param.print_diagnostics()
print('Prior logprob: %f' % param.prior_logprob())
raise Exception("Prior returned %f logprob" % lp)
if not np.isfinite(lp):
return lp
# include the log probability from the model
lp += model.log_likelihood()
if np.isnan(lp):
raise Exception("Likelihood returned %f logprob" % lp)
return lp
def sample(self, model):
for i in range(self.thinning + 1):
params_array, new_latent_values, current_ll = self.sample_fun(
model, **self.sampler_options)
hyperparameter_utils.set_params_from_array(self.params,
params_array)
model.latent_values.set_value(new_latent_values)
self.current_ll = current_ll # for diagnostics
def logprob(self, x, model, nu):
hyperparameter_utils.set_params_from_array(self.params, x)
lp = 0.0
for param in self.params:
lp += param.prior_logprob()
# Compute this if prior logprob is finite AND there is data
if np.isfinite(lp) and nu is not None:
# Get implied y from nu
implied_y = self._compute_implied_y(model, nu)
# Instead of marginal likelihood, check binomial likelihood
lp += model.log_binomial_likelihood(y=implied_y)
return lp
def sample(self, model):
if not model.has_data:
return np.zeros(
0) # TODO this should be a sample from the prior...
prior_cov = model.noiseless_kernel.cov(model.inputs)
prior_cov_chol = spla.cholesky(prior_cov, lower=True)
# Here get the Cholesky from model
params_array = hyperparameter_utils.params_to_array(self.params)
for i in range(self.thinning + 1):
params_array, current_ll = elliptical_slice(
params_array, self.logprob, prior_cov_chol, model.mean.value,
model, **self.sampler_options)
hyperparameter_utils.set_params_from_array(self.params,
params_array)
self.current_ll = current_ll # for diagnostics
def sample(self, model):
"""generate a new sample of parameters for the model
Notes
-----
The parameters are stored as self.params which is a list of Params objects.
The values of the parameters are updated on each call. Pesumably the value of
the parameter affects the model (this is not required, but it would be a bit
pointless othewise)
"""
# turn self.params into a 1d numpy array
params_array = hyperparameter_utils.params_to_array(self.params)
for i in range(self.thinning + 1):
# get a new value for the parameter array via slice sampling
params_array, current_ll = slice_sample(params_array, self.logprob,
model,
**self.sampler_options)
hyperparameter_utils.set_params_from_array(
self.params, params_array) # Can this be untabbed safely?
self.current_ll = current_ll # for diagnostics
def logprob(self, x, model):
hyperparameter_utils.set_params_from_array(self.params, x)
return model.log_binomial_likelihood(
) # no contribution from priors-- Gaussian prior built in to sampler