Deterministic AE Family

Base tree

BaseAutoencoderConfig
├── latent_dim
└── reconstruction_loss

Models

AutoencoderConfig
└── no additional model fields

SemanticHashingAutoencoderConfig
├── binarization
├── binarization_weight
├── balance_weight
└── decorrelation_weight

DenoisingAutoencoderConfig
├── noise_type
├── noise_std
└── masking_ratio

ContractiveAutoencoderConfig
└── contractive_weight

SparseAutoencoderConfig
├── sparsity_weight
└── target_activation

TopKSparseAutoencoderConfig
└── topk

KLSparseAutoencoderConfig
├── sparsity_weight
└── target_activation

WassersteinAutoencoderConfig
├── mmd_weight
└── kernel_bandwidths

AdversarialAutoencoderConfig
├── adversarial_weight
└── discriminator_hidden_dims

Notes

• Deterministic families usually allow decoder: null when reversing the encoder preserves the decoder runtime input spec.
• Architecture-specific fields belong in encoder and decoder module configs rather than here.
• SemanticHashingAutoencoder keeps the same single-sample TensorSpec contract as ordinary autoencoders, so it fits the existing AETrainer and backbone modules without requiring a pairwise dataset format.