def fvp(self, xs, g, **kwargs):
""" Return the product between a vector g (in the same formast as
self.variable) and the Fisher information defined by the average
over xs. """
gs = unflatten(g, shapes=self.var_shapes)
ts_fvp = self.ts_fvp(array_to_ts(xs), array_to_ts(gs), **kwargs)
return flatten([v.numpy() for v in ts_fvp])
def fvp0(self, xs, ys, g, **kwargs):
""" Computes F(self.pi)*g, where F is the Fisher information matrix and
g is a np.ndarray in the same shape as self.variable, were the Fisher
information defined by the average over xs. """
gs = unflatten(g, shapes=self.var_shapes)
ts_fvp = self.ts_fvp0(array_to_ts(xs), array_to_ts(ys), array_to_ts(gs), **kwargs)
return flatten([v.numpy() for v in ts_fvp])
def logp_grad(self, xs, ys, fs, **kwargs):
ts_grad = self.ts_logp_grad(array_to_ts(xs), array_to_ts(ys),
array_to_ts(fs), **kwargs)
return flatten([v.numpy() for v in ts_grad])
def exp_grad(self, xs, As, bs, cs, canonical=True, diagonal_A=True):
""" See exp_fun. """
ts_grad = self.ts_exp_grad(array_to_ts(xs), array_to_ts(As),
array_to_ts(bs), array_to_ts(cs),
canonical, diagonal_A)
return flatten([v.numpy() for v in ts_grad])
def mean_variable(self):
return flatten(ts_to_array(super().ts_variables))
def grad(self, x, **kwargs):
return flatten(ts_to_array(self.ts_grad(array_to_ts(x), **kwargs)))
def flatten(self, vs):
return flatten(vs)
def compute_grad(self):
if self._args is None:
raise ValueError('Oracle has not been initialized')
grads = self._compute_grad(*self._args)
return flatten(grads)
def flatten(self, vals):
return flatten(vals)