real_nvp module

Implementation of realNVP: a coupling cell with an affine transform

class ElementWiseAffineTransform(*args, **kwargs)[source]

Bases: zunis.models.flows.coupling_cells.transforms.InvertibleTransform

Invertible element-wise affine transform

Constructor for InvertibleTransform

No arguments, no returns, this is just encapsulation for a function and its inverse

backward(y, st, compute_jacobian=True)[source]
forward(y, st, compute_jacobian=True)[source]
class FakeRealNVP(*args, **kwargs)[source]

Bases: zunis.models.flows.coupling_cells.real_nvp.GeneralRealNVP

Fake real NVP with position independent affine parameters for basic debugging and training tests

Parameters

  • d (int) – dimension of the space

  • transform (InvertibleTransform) – bijective variable transform that maps the unit hypercube (in d-sum(mask) dimensions) to itself given some parameters.

  • mask (list of bool) – indicates which variables are passed (y_N) through and which are transformed (y_M)

training: bool[source]
class FakeT(s=0.0, t=0.0)[source]

Bases: torch.nn.modules.module.Module

Fake neural net for the test FakeRealNVP that returns a trainable constant

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x)[source]
training: bool[source]
class GeneralRealNVP(*args, **kwargs)[source]

Bases: zunis.models.flows.coupling_cells.general_coupling.InvertibleCouplingCell

Coupling cell based on affine transforms without a specific parameter prediction function

Parameters

  • d (int) – dimension of the space

  • transform (InvertibleTransform) – bijective variable transform that maps the unit hypercube (in d-sum(mask) dimensions) to itself given some parameters.

  • mask (list of bool) – indicates which variables are passed (y_N) through and which are transformed (y_M)

training: bool[source]
class RealNVP(*args, **kwargs)[source]

Bases: zunis.models.flows.coupling_cells.real_nvp.GeneralRealNVP

Real non-volume-preserving coupling cell.

This implements a bijective mapping R^d -> R^d by transforming some subset of the coordinates defined by the mask through a coordinate-wise affine transformation. As with other coupling cells the parameters of the affine transformation are the output of a neural network taking the un-tranformed coordinates as input.

Parameters

  • d (int) – dimension of the space

  • mask (list of bool) – mask defining which variables are transformed

  • d_hidden (int, 256) – width of the neural network

  • n_hidden (int, 8) – depth of the neural network

  • input_activation – activation function applied before going through the neural network

  • hidden_activation – activation function applied after each hidden layer of the neural network

  • output_activation – activation function applied after the last layer of the neural network

  • use_batch_norm (bool) – whether to use batch normalization after each hidden layer of the neural network

training: bool[source]
element_wise_affine(x, st, compute_jacobian=True)[source]

Transform x element-wise through an affine function y = exp(s)*x + t where s = st[…,0] and t = st[…,1] with s.shape == x.shape == t.shape

The Jacobian for this transformation is the coordinate-wise product of the scaling factors J = prod(es[…,i],i)

inverse_element_wise_affine(x, st, compute_jacobian=True)[source]

Transform x element-wise through an affine function y = exp(-s)*(x - t) where s = st[…,0] and t = st[…,1] with s.shape == x.shape == t.shape This is the inverse of element_wise_affine above for the same set of parameters st

The Jacobian for this transformation is the coordinate-wise product of the scaling factors J = prod(es[…,i],i)