NICEConfig

Configuration for every NICE variant.

Defaults reproduce the paper's MNIST experiment (Figure 3): 784 dimensions, four coupling layers whose coupling networks have five hidden layers of 1000 rectified units, and a standard logistic prior.

Args: input_dim: Dimensionality D of the data the flow is defined on — the paper's "# dimensions" row and the released configs' nvis. Figure 3: 784 (MNIST), 2304 (TFD), 3072 (SVHN, CIFAR-10). Must be even, since every coupling layer splits it in half. num_coupling_layers: Number of stacked additive coupling layers. The paper uses 4 for every dataset; three is the minimum that lets all dimensions interact. num_hidden_layers: Hidden layers inside each coupling network m. Paper Figure 3: 5 for MNIST, 4 for TFD / SVHN / CIFAR-10. hidden_dim: Units per hidden layer of m. Paper Figure 3: 1000 (MNIST), 5000 (TFD), 2000 (SVHN, CIFAR-10). input_reorder: One-off permutation applied before the first coupling layer. "tile" (default) reproduces the paper's odd/even partition; "none" splits in raster order, which is what the released configs for SVHN / CIFAR-10 do. init_range: Half-width of the uniform weight initialisation of every coupling network, matching the official release's irange = 0.01. Biases start at zero, and the diagonal scaling starts at s = 0 (identity).

Notes: NICE is defined on flat D-vectors, not on a pixel grid: both the paper and the released configs parametrise each experiment by a single dimension count. input_dim is therefore the only dimensional knob, and the spatial fields inherited from GenerativeModelConfig (sample_size, in_channels, out_channels) are pinned to the equivalent vector encoding rather than to an image shape the sources never state.