TSNPE spends its whole budget near one observation, so a mis-sized x_obs
is not a detail. Without this check as_theta_matrix() folds a length-3
vector into a 3 x 1 matrix and resolve_x() keeps row 1, so the run
truncates its proposals around a value the user never asked for, with only a
warning to go on. Called twice: once at the top of npe_sequential() for
shape, and once after round 1 with dim_x, which is the first moment the
simulator's output width is known.
Arguments
- x_obs
The observation passed to
npe_sequential().- dim_x
Simulator output width, or
NULLto check shape only.