def __init__(self, neg_log_dens, grad_neg_log_dens=None):
"""
Args:
neg_log_dens (Callable[[array], float]): Function which given a
position array returns the negative logarithm of an
unnormalised probability density on the position space with
respect to the Lebesgue measure, with the corresponding
distribution on the position space being the target
distribution it is wished to draw approximate samples from.
grad_neg_log_dens (
None or Callable[[array], array or Tuple[array, float]]):
Function which given a position array returns the derivative of
the negative logarithm of the unnormalised density specified by
`neg_log_dens` with respect to the position array argument.
Optionally the function may instead return a pair of values
with the first being the array corresponding to the derivative
and the second being the value of the `neg_log_dens` evaluated
at the passed position array. If `None` is passed (the default)
an automatic differentiation fallback will be used to attempt
to construct the derivative of `neg_log_dens` automatically.
"""
self._neg_log_dens = neg_log_dens
self._grad_neg_log_dens = autodiff_fallback(grad_neg_log_dens,
neg_log_dens,
'grad_and_value',
'grad_neg_log_dens')
def __init__(self,
neg_log_dens,
grad_neg_log_dens=None,
hess_neg_log_dens=None,
mtp_neg_log_dens=None,
softabs_coeff=1.):
self._hess_neg_log_dens = autodiff_fallback(hess_neg_log_dens,
neg_log_dens,
'hessian_grad_and_value',
'neg_log_dens')
self._mtp_neg_log_dens = autodiff_fallback(
mtp_neg_log_dens, neg_log_dens, 'mtp_hessian_grad_and_value',
'mtp_neg_log_dens')
super().__init__(neg_log_dens,
SoftAbsRegularisedPositiveDefiniteMatrix,
self._hess_neg_log_dens,
self._mtp_neg_log_dens,
grad_neg_log_dens,
metric_kwargs={'softabs_coeff': softabs_coeff})
def __init__(self,
neg_log_dens,
metric_matrix_class,
metric_func,
vjp_metric_func=None,
grad_neg_log_dens=None,
metric_kwargs=None):
self._metric_matrix_class = metric_matrix_class
self._metric_func = metric_func
self._vjp_metric_func = autodiff_fallback(vjp_metric_func, metric_func,
'vjp_and_value',
'vjp_metric_func')
self._metric_kwargs = {} if metric_kwargs is None else metric_kwargs
super().__init__(neg_log_dens, grad_neg_log_dens)
def __init__(self,
neg_log_dens,
constr,
metric=None,
dens_wrt_hausdorff=True,
grad_neg_log_dens=None,
jacob_constr=None):
super().__init__(neg_log_dens=neg_log_dens,
metric=metric,
grad_neg_log_dens=grad_neg_log_dens)
self._constr = constr
self.dens_wrt_hausdorff = dens_wrt_hausdorff
self._jacob_constr = autodiff_fallback(jacob_constr, constr,
'jacobian_and_value',
'jacob_constr')
def __init__(self,
neg_log_dens,
constr,
metric=None,
dens_wrt_hausdorff=True,
grad_neg_log_dens=None,
jacob_constr=None,
mhp_constr=None):
super().__init__(neg_log_dens=neg_log_dens,
constr=constr,
metric=metric,
dens_wrt_hausdorff=dens_wrt_hausdorff,
grad_neg_log_dens=grad_neg_log_dens,
jacob_constr=jacob_constr)
if not dens_wrt_hausdorff:
self._mhp_constr = autodiff_fallback(mhp_constr, constr,
'mhp_jacobian_and_value',
'mhp_constr')
