ncsn_celeba

→NCSNModel

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

source edit

Construct an NCSN score network for the CelebA 64x64 setup.

Paper-faithful CelebA 64x64 configuration from NCSNv2 (Song 2020) Appendix C: sample size 64x64, base_channels=128, channel_mult=(1, 1, 2, 2, 4, 4), 2 ResBlocks per stage, self-attention at the 16x16 feature map, $L = 500$ geometric noise levels in $[\sigma_{\min} = 0.01, \sigma_{\max} = 90]$ , and the NCSNv2 Langevin schedule ( $T = 5$ steps per level, $\epsilon = 2 \cdot 10^{-5}$ ).

Model Size

Parameters

pretrainedbool= False

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

**overridesobject= {}

Optional NCSNConfig field overrides forwarded into the underlying config.

Returns

NCSNModel

Score network configured with the CelebA 64x64 setup and any overrides.

Notes

Reference: Song and Ermon, "Improved Techniques for Training Score-Based Generative Models" (NCSNv2), NeurIPS, 2020 (arXiv:2006.09011), Appendix C.

Examples

>>> import lucid
>>> from lucid.models.generative.ncsn import ncsn_celeba
>>> model = ncsn_celeba().eval()
>>> x_tilde = lucid.randn((1, 3, 64, 64))
>>> sigma_idx = lucid.tensor([10]).long()
>>> out = model(x_tilde, sigma_idx)
>>> out.sample.shape   # (1, 3, 64, 64)
(1, 3, 64, 64)

Used by 1

lucid.models.generative.ncsn

>>> import lucid >>> from lucid.models.generative.ncsn import ncsn_celeba >>> model = ncsn_celeba().eval() >>> x_tilde = lucid.randn((1, 3, 64, 64)) >>> sigma_idx = lucid.tensor([10]).long() >>> out = model(x_tilde, sigma_idx) >>> out.sample.shape # (1, 3, 64, 64) (1, 3, 64, 64)