v2.0.1

InvertibleNetworks v2.0.1

Diff since v2.0.0

Closed issues:

• JOLI (#25)
• GPU support of InvertibleNetworks in julia 1.6.1 (#30)

Merged pull requests:

• Typos (#28) (@PallHaraldsson)
• add SLIM registry (#29) (@alisiahkoohi)
• Hyperbolic network now uses GPU supported Haar wavelet (#31) (@PetersBas)
• add compat for registration (#33) (@mloubout)

Chainrules

• Support for chainrules and Zygote
• Fix GPU support
• Make layers/network size independent to follow general conventions and generalizations.

GPU support

Fixes GPU support:

• Fix dimensionality manipulation views
• Fix loss function input types
• Add GPU supported Haar wavelet squeeze/unsqueeze

Simplifies and fix bugs in Hyperbolic layer
Add documentation

Jacobians and adjoint Jacobians of layers and networks

Implementation of Jacobian evaluation (and adjoint thereof) for invertible networks

Main contributions:

• added Jacobian functionalities for most of the layers and networks already implemented, except for layers and networks requiring a "link" function (invertible_layer_slim_additive, invertible_layer_slim_affine, invertible_network_irim) and Flux blocks that requires forward-mode AD;

• Jacobians can be wrapped around an abstract linear algebra framework, see /examples/jacobian/jacobian_linear_algebra.jl;

• composition functionalities to string networks/layers together, see /examples/utils/composition_of_networks.jl
  Jacobians also work together with Composition.

• new layers/networks: layer_resnet (as a Flux block), invertible_network_hint_multiscale (previously only had a conditional version)

• utilities for Parameter or Array{Parameter, 1} direct manipulation, see /examples/utils/parameter_algebra.jl

v1.0.1b

Support of Jacobians and adjoint of layers and networks.

First release

v.1.0.0

double assignment