ncsn_cifar

→NCSNModel

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

source edit

Construct an NCSN score network for the CIFAR-10 setup.

Paper-faithful CIFAR-10 configuration combining Song and Ermon, 2019 §4.1 with the NCSNv2 (Song 2020) Table 1 refinements: sample size 32x32, base_channels=128, channel_mult=(1, 2, 2, 2), 2 ResBlocks per stage, self-attention at the 16x16 feature map, $L = 232$ geometric noise levels in $[\sigma_{\min} = 0.01, \sigma_{\max} = 50]$ , and the NCSNv2 §3.3 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 CIFAR-10 setup and any overrides.

Notes

Reference: Song and Ermon, "Generative Modeling by Estimating Gradients of the Data Distribution", NeurIPS, 2019 (arXiv:1907.05600); NCSNv2 refinements in Song and Ermon, 2020 (arXiv:2006.09011).

Examples

>>> import lucid
>>> from lucid.models.generative.ncsn import ncsn_cifar
>>> model = ncsn_cifar().eval()
>>> x_tilde = lucid.randn((1, 3, 32, 32))
>>> sigma_idx = lucid.tensor([3]).long()
>>> out = model(x_tilde, sigma_idx)
>>> out.sample.shape   # (1, 3, 32, 32) — raw score
(1, 3, 32, 32)

Used by 1

lucid.models.generative.ncsn

>>> import lucid >>> from lucid.models.generative.ncsn import ncsn_cifar >>> model = ncsn_cifar().eval() >>> x_tilde = lucid.randn((1, 3, 32, 32)) >>> sigma_idx = lucid.tensor([3]).long() >>> out = model(x_tilde, sigma_idx) >>> out.sample.shape # (1, 3, 32, 32) — raw score (1, 3, 32, 32)