nice_cifar

→NICEModel72.6M

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

source edit

Construct the NICE flow for the CIFAR-10 setup.

Paper-faithful CIFAR-10 configuration from Dinh et al., 2014 Figure 3: 3072 dimensions, four coupling layers with four hidden layers of 2000 units each, and a standard logistic prior.

Model Size

Parameters

pretrainedbool= False

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

**overridesobject= {}

Optional NICEConfig field overrides forwarded into the underlying config.

Returns

NICEModel

Bare bijection configured with the CIFAR-10 setup and any overrides.

Notes

Reference: Dinh, Krueger, and Bengio, "NICE: Non-linear Independent Components Estimation", ICLR Workshop, 2015 (arXiv:1410.8516), Figure 3. Reported test log-likelihood: 5371.78 nats, on ZCA-whitened data dequantised with uniform 1/128 noise and rescaled to $[-1, 1]^D$ (CIFAR-10 is the one dataset given the wider range). The released reference config for this dataset instead stacks 8 coupling layers of 2400 units with a normal prior and no input reordering — see the module docstring for how far overrides reproduce it.

Examples

>>> import lucid
>>> from lucid.models.generative.nice import nice_cifar
>>> model = nice_cifar(input_dim=64, hidden_dim=16).eval()
>>> x = lucid.rand((1, 64)) * 2.0 - 1.0
>>> model.log_prob(x).shape
(1,)

Used by 1

lucid.models.generative.nice

>>> import lucid >>> from lucid.models.generative.nice import nice_cifar >>> model = nice_cifar(input_dim=64, hidden_dim=16).eval() >>> x = lucid.rand((1, 64)) * 2.0 - 1.0 >>> model.log_prob(x).shape (1,)