vqvae_gen

→VQVAEForImageGeneration

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

source edit

Construct a VQ-VAE with the full training objective and a sampler.

Same trunk as vqvae, wrapped with the three-term objective of van den Oord et al., 2017 — reconstruction, codebook, and $\beta$ -weighted commitment — plus a convenience sampler over the uniform codebook prior.

Model Size

Parameters

pretrainedbool= False

No weights are published for this family; passing True raises rather than returning a randomly initialised model.

**overridesobject= {}

Optional VQVAEConfig field overrides forwarded into the underlying config. Pass recon_loss="bce" for Bernoulli likelihoods on [0, 1] data, or commitment_cost=... to retune

\beta

Returns

VQVAEForImageGeneration

VQ-VAE wrapped with the full objective and prior sampler.

Notes

Reference: van den Oord, Vinyals, and Kavukcuoglu, "Neural Discrete Representation Learning", NeurIPS, 2017 (arXiv:1711.00937).

Training objective:

L = \log p(x \mid z_q(x)) + \big\| \mathrm{sg}[z_e(x)] - e \big\|_2^2 + \beta \big\| z_e(x) - \mathrm{sg}[e] \big\|_2^2 .

generate samples the uniform prior the model was trained against, not a learned one — the paper fits a PixelCNN over the latent grid for its figures. Samples from the uniform prior are expected to be incoherent; see VQVAEForImageGeneration.

Examples

>>> import lucid
>>> from lucid.models.generative.vqvae import vqvae_gen
>>> model = vqvae_gen().eval()
>>> x = lucid.randn((1, 3, 32, 32))
>>> out = model(x)
>>> out.sample.shape, out.perplexity.shape
((1, 3, 32, 32), ())

Used by 1

lucid.models.generative.vqvae

>>> import lucid >>> from lucid.models.generative.vqvae import vqvae_gen >>> model = vqvae_gen().eval() >>> x = lucid.randn((1, 3, 32, 32)) >>> out = model(x) >>> out.sample.shape, out.perplexity.shape ((1, 3, 32, 32), ())