NVPFlow

class NVPFlow(features, hidden_features, num_layers, num_blocks_per_layer, use_volume_preserving=False, activation=<function relu>, dropout_probability=0.0, batch_norm_within_layers=False, batch_norm_between_layers=False, random_permutation=True)

Bases: Flow

A simplified version of Real NVP for 1-dim inputs.

This implementation uses 1-dim checkerboard masking but doesn’t use multi-scaling.

Reference: L. Dinh et al., Density estimation using Real NVP, ICLR 2017.

This class has been modified from the example found at: https://github.com/bayesiains/nflows/blob/master/nflows/flows/realnvp.py

__init__(features, hidden_features, num_layers, num_blocks_per_layer, use_volume_preserving=False, activation=<function relu>, dropout_probability=0.0, batch_norm_within_layers=False, batch_norm_between_layers=False, random_permutation=True)

Parameters:

features: int

The number of input dimensions.

hidden_features: int

The number of hidden features in each residual network.

num_layers: int

The number of coupling layers.

num_blocks_per_layer: int

The number of residual blocks in each coupling layer.

use_volume_preserving: bool

Whether to use additive (volume-preserving) rather than affine coupling transforms.

activation: callable

The activation function for the residual networks.

dropout_probability: float

The dropout probability in the residual networks.

batch_norm_within_layers: bool

Whether to use batch normalization within the residual networks.

batch_norm_between_layers: bool

Whether to apply batch normalization between coupling layers.

random_permutation: bool

Whether to randomly permute the inputs between coupling layers.

__call__(*args, **kwargs)

Call self as a function.

Methods

__init__(features, hidden_features, ...[, ...])

Attributes

call_super_init
dump_patches
training