general_backprop_j_flow module¶
This module implements an abstract class for flows defined as the application of a differentiable pytorch function on an input point. In train mode, the last entry of the input is the logarithm of the inverse PDF at that point.
- class GeneralBackpropJacobianFlow(*args, **kwargs)[source]¶
Bases:
zunis.models.flows.general_flow.GeneralFlow
Abstract class for a flow defined as a general differentiable Pytorch function This implements the universal transform_and_compute_jacobian behavior: y = flow(x), - log(j(y)) = - log(j(x)) + log(det(dy_i/dx_j)) The jacobian is computed naively by using autograd.
Initializes internal Module state, shared by both nn.Module and ScriptModule.