Construct the NICE flow for the Toronto Face Database setup.
Paper-faithful TFD configuration from Dinh et al., 2014 Figure 3: 2304 dimensions, four coupling layers with four hidden layers of 5000 units each, and — uniquely among the four experiments — a standard normal prior.
Model Size
Parameters
pretrainedbool= FalseReserved for future weight registration; currently a no-op.
**overridesobject= {}Optional
NICEConfig field overrides forwarded into the
underlying config.Returns
NICEModelBare bijection configured with the TFD 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: 5514.71 nats, on approximately whitened data — well ahead of the 5250 nats of the Deep MFA bound it was compared against.
Examples
>>> import lucid
>>> from lucid.models.generative.nice import nice_tfd
>>> model = nice_tfd(input_dim=64, hidden_dim=16).eval()
>>> model.prior
'gaussian'
>>> model.input_dim
64