def _precision(self):
prec = torch.diag(self.cm) / self.cm.sum(dim=0)
prec[prec != prec] = 0 # removing potential "nan"s
if self.average is None: return prec
else:
weights = self._weights(avg=self.average)
return (prec * weights).sum()
def _recall(self):
rec = torch.diag(self.cm) / self.cm.sum(dim=1)
rec[rec != rec] = 0 # removing potential "nan"s
if self.average is None: return rec
else:
if self.average == "micro": weights = self._weights(avg="weighted")
else: weights = self._weights(avg=self.average)
return (rec * weights).sum()
def _get_style_vals(self, tensor):
mean, cov = self._calc_2_moments(tensor)
if mean is None:
return None, None, None
eigvals, eigvects = torch.symeig(cov, eigenvectors=True)
eigroot_mat = torch.diag(torch.sqrt(eigvals.clamp(min=0)))
root_cov = torch.mm(torch.mm(eigvects, eigroot_mat), eigvects.t())
tr_cov = eigvals.clamp(min=0).sum()
return mean, tr_cov, root_cov