realnvp_celeba_gen(pretrained: bool = False, overrides: object = {})RealNVP generator for CelebA 64 x 64.
Same bijection as realnvp_celeba, wrapped with the bits/dim
objective and the ancestral sampler. This is the setup the paper's
latent-space interpolation figures come from — the flow is a
bijection, so interpolating between two encode outputs and
decoding gives a semantically smooth path with no encoder
approximation.
Model Size
Parameters
pretrainedbool= FalseReserved for future weight registration; currently a no-op.
**overridesobject= {}Optional
RealNVPConfig field overrides forwarded into
the underlying config.Returns
RealNVPForImageGenerationGenerator configured with the CelebA setup.
Notes
Reference: Dinh, Sohl-Dickstein, and Bengio, "Density Estimation Using Real NVP", ICLR, 2017 (arXiv:1605.08803). Reported: 3.02 bits/dim.
Examples
>>> from lucid.models.generative.realnvp import realnvp_celeba_gen
>>> model = realnvp_celeba_gen().eval()
>>> model.generate(n_samples=1).samples.shape
(1, 3, 64, 64)