def utility_term_Jensen(us, data_mask):
#print("[utility_term_Jensen] us=%s" % us.shape)
us = flatten_first_two_dims(us)
point_utility_term = (us + EPS_JENSEN).log().mean(0)
assert point_utility_term.shape == data_mask.shape, "%s=datashape!=mask.shape=%s" % (
point_utility_term.shape, data_mask.shape)
return torch.masked_select(point_utility_term, data_mask).sum()
def sample_predictive_y(qw, qz, nsamples_theta, nsamples_y):
""" Returns a tensor with samples (nsamples_y x nsamples_theta).
Flattents the first two dimensions
(samples of y for different thetas) from sample_predictive_y0.
"""
return flatten_first_two_dims(
sample_predictive_y0(qw, qz, nsamples_theta, nsamples_y))
def optimize_h_with_bayes_estimator(*args):
"""Assumes that optimal_h calculates gain-optimal value analytically."""
ys = flatten_first_two_dims(
self.sample_predictive_y0(
*args,
nsamples_theta=self.H_NSAMPLES_UTILITY_TERM_THETA,
nsamples_y=self.H_NSAMPLES_UTILITY_TERM_Y))
h = self.optimal_h_bayes_estimator(ys)
if optimize_h_with_bayes_estimator.counter < 3 and not is_valid(h):
print(
"WARNING: your optimal_h_bayes_estimator=%s returns invalid decisions! Are we good here?"
% self.optimal_h_bayes_estimator.__name__)
optimize_h_with_bayes_estimator.counter += 1
return h