def mutual_info_gap(
self,
n_bins: int = 10,
strategy: Literal['uniform', 'quantile', 'kmeans', 'gmm'] = 'uniform',
) -> float:
"""Mutual Information Gap
Parameters
----------
n_bins : int, optional
number of bins for discretizing the latents, by default 10
strategy : {'uniform', 'quantile', 'kmeans', 'gmm'}
Strategy used to define the widths of the bins.
'uniform' - All bins in each feature have identical widths.
'quantile' - All bins in each feature have the same number of points.
'kmeans' - Values in each bin have the same nearest center of a 1D cluster.
, by default 'uniform'
Returns
-------
float
mutual information gap score
"""
z = self.dist_to_tensor(self.latents).numpy()
if n_bins > 1:
z = discretizing(z,
independent=True,
n_bins=n_bins,
strategy=strategy)
f = self.factors
return mutual_info_gap(z, f)
def cal_mutual_info_gap(self, mean=True):
r"""
Arguments:
mean : a Boolean, if True use the mean of latent distribution for
calculating the mutual information gap
Return:
a dictionary : {'mig': score}
Reference:
Chen, R.T.Q., Li, X., Grosse, R., Duvenaud, D., 2019. Isolating Sources of
Disentanglement in Variational Autoencoders. arXiv:1802.04942 [cs, stat].
"""
mig = []
for z, f in zip(self._latent_codes(mean), self.factors):
mig.append(metrics.mutual_info_gap(z, f))
return dict(mig=np.mean(mig))