hvae

→VAEModel

hvae(pretrained: bool = False, overrides: object = {})

source edit

Construct a hierarchical (Ladder-style) VAE with one latent per stage.

Hierarchical VAE topology following Sønderby et al., 2016, configured with a 32x32 input, down_block_channels=(64, 128, 256) (three stride-2 encoder stages), and per-stage latent widths (32, 64, 128) — one $z_l \in \mathbb{R}^{d_l}$ extracted per stage and re-injected at the matching decoder resolution. The KL becomes a sum over levels (independent-posterior Ladder variant).

Model Size

Parameters

pretrainedbool= False

Reserved for future weight registration; currently a no-op.

**overridesobject= {}

Optional VAEConfig field overrides forwarded into the underlying config. The tuple latent_dim must match the length of down_block_channels.

Returns

VAEModel

Hierarchical VAE trunk configured with the 3-level default and any overrides.

Notes

Reference: Sønderby, Raiko, Maaløe, Sønderby, and Winther, "Ladder Variational Autoencoders", NeurIPS, 2016 (arXiv:1602.02282).

Hierarchical ELBO:

\mathcal{L}_{\mathrm{ELBO}} = \mathbb{E}_{q_\phi(z \mid x)}\!\big[\log p_\theta(x \mid z)\big] - \sum_{l = 1}^{L} \mathrm{KL}\!\big( q_\phi(z_l \mid x) \,\big\|\, p(z_l) \big).

Examples

>>> import lucid
>>> from lucid.models.generative.vae import hvae
>>> model = hvae().eval()
>>> x = lucid.randn((1, 3, 32, 32))
>>> out = model(x)
>>> out.sample.shape, out.latent.shape   # latent is flat concat (32+64+128)
((1, 3, 32, 32), (1, 224))

Used by 1

lucid.models.generative.vae

>>> import lucid >>> from lucid.models.generative.vae import hvae >>> model = hvae().eval() >>> x = lucid.randn((1, 3, 32, 32)) >>> out = model(x) >>> out.sample.shape, out.latent.shape # latent is flat concat (32+64+128) ((1, 3, 32, 32), (1, 224))