realnvp_celeba(pretrained: bool = False, overrides: object = {})Construct the RealNVP flow for the CelebA 64 x 64 setup.
Paper-faithful configuration from Dinh et al., 2016 §4.1 — five
scales with 2-residual-block coupling networks, matching the other
64 x 64 experiments.
Model Size
Parameters
pretrainedbool= FalseNo checkpoint is published for
realnvp_celeba, so True raises
NotImplementedError rather than return randomly initialised weights.**overridesobject= {}Optional
RealNVPConfig field overrides forwarded into
the underlying config.Returns
RealNVPModelBare bijection configured with the CelebA setup.
Notes
Reference: Dinh, Sohl-Dickstein, and Bengio, "Density Estimation
Using Real NVP", ICLR, 2017 (arXiv:1605.08803), §4.1 and Table 1.
Reported validation result: 3.02 bits/dim (2.97 train), on an
approximately central 148 x 148 crop resized to 64 x 64.
Examples
>>> from lucid.models.generative.realnvp import realnvp_celeba
>>> model = realnvp_celeba().eval()
>>> model.prior
'gaussian'