VariationalEncoder

latent.modules.encoder.VariationalEncoder

Variational encoder. This model compresses input data by parameterizing a latent distribution that is regularized through a KL Divergence loss.

__init__(self, latent_dim=50, name='variational_encoder', initializer='glorot_normal', use_decomposed_kld=False, x_size=1000, kld_weight=0.0001, tc_weight=0.001, capacity=0.0, prior='normal', latent_dist='independent', iaf_units=[256, 256], n_pseudoinputs=200, **kwargs) special

Parameters:

Name	Type	Description	Default
latent_dim	int	Integer indicating the number of dimensions in the latent space.	50
name	str	String indicating the name of the model.	'variational_encoder'
initializer	Union[str, Callable]	Initializer for the kernel weights matrix (see keras.initializers)	'glorot_normal'
use_decomposed_kld	bool	Boolean indicating whether to use the decomposed KLD loss (Chen 2019)	False
x_size	int	Total number of data points. Only used if use_decomposed_kld = True.	1000
kld_weight	float	Float indicating the weight of the KL Divergence regularization loss. If use_decomposed_kld = True, this indicated the weight of the dimension-wise KLD.	0.0001
tc_weight	float	Float indicating the weight of the total correlation term of the KLD loss. Only used if use_decomposed_kld = True.	0.001
capacity	float	Capacity of the KLD loss. Can be linearly increased using a KL scheduler callback.	0.0
prior	Literal['normal', 'iaf', 'vamp']	The choice of prior distribution. One of the following: 'normal' - A unit gaussian (normal) distribution. 'iaf' - A unit gaussian with a Inverse Autoregressive Flows bijector (Kingma 2016) 'vamp' - A variational mixture of posteriors (VAMP) prior (Tomczak 2017)	'normal'
latent_dist	Literal['independent', 'multivariate']	The choice of latent distribution. One of the following: 'independent' - A independent normal produced by tfpl.IndependentNormal. 'multivariate' - A multivariate normal produced by tfpl.MultivariateNormalTriL.	'independent'
iaf_units	Iterable[int]	Integer list indicating the units in the IAF bijector network. Only used if prior = 'iaf'.	[256, 256]
n_pseudoinputs	int	Integer indicating the number of pseudoinputs for the VAMP prior. Only used if prior = 'vamp'.	200
**kwargs		Other arguments passed on to DenseStack.	{}