Variational Family

Base tree

BaseVariationalAutoencoderConfig
├── BaseAutoencoderConfig
│   ├── latent_dim
│   └── reconstruction_loss
├── kl_weight
├── free_bits
├── kl_warmup_epochs
├── kl_start_weight
└── use_mean_in_eval

Models

VariationalAutoencoderConfig
└── no additional model fields

BetaVariationalAutoencoderConfig
└── inherits base variational fields; beta behavior is expressed through kl_weight

DenoisingVariationalAutoencoderConfig
├── noise_type
├── noise_std
└── masking_ratio

HierarchicalVariationalAutoencoderConfig
└── top_latent_dim

VampPriorVariationalAutoencoderConfig
└── num_pseudo_inputs

InformationVariationalAutoencoderConfig
├── mmd_weight
└── kernel_bandwidths

MMDVariationalAutoencoderConfig
├── mmd_weight
└── kernel_bandwidths

DIPVariationalAutoencoderConfig
├── dip_type
├── lambda_diag
└── lambda_offdiag

BetaTCVariationalAutoencoderConfig
└── tc_weight

FactorVariationalAutoencoderConfig
├── tc_weight
└── discriminator_hidden_dims

Notes

• Variational models often change decoder runtime input space from encoder output space.
• Because of that, decoder: null is intentionally rejected for many VAE-family runs unless the specs still match exactly.