Complex-valued neural networks for pytorch and Variational Dropout for real and complex layers.
In 2022.06 major release we increases the minimal versions to python>=3.7 and pytorch>=1.8. Although the modern torch now natively supports complex dtypes, no transition has been made to use them as the new backend, and currently we still use the split representation with CR-calculus on top (see the discussions in issues #2 and #21 ).
The following features have been added:
3.7 as pointed out in issue #24.utils.spectrum to new native torch complex backend (torch>=1.8).Cplx instances deepcopy-able, fixing issue #18The following cosmetic or repo-level modifications have been made:
.nn.ModReLU indicating the sign-deviation from the original paper proposing it (issue #22)This is a nominal major release, as it increases the minimal pytorch version from 1.4 to 1.7.
The following features have been added:
The version has been bumped from 2020 to 2021 to reflect the new year.
This is a minor mid-month release.
The following features were added:
Cplx # 7, and support for view and view_as methods for Cplx # 6 by Hendrik Schröter
The following bugs were fixed:
nn.init.cplx_trabelsi_independent_, which prevented it from working properly # 11
This is a minor release, that adds support for 3d real- and complex-valued convolutions and Variational Dropout for them.
This release includes a fix that makes masked layers work in multi gpu setting, and a update to sparsity accounting.
An extension for torch that adds basic building blocks for complex-valued neural networks with batch normalization and weight initialization. Provides an implementation of Real- and Complex-valued Bayesian sparisification techniques: Variational Dropout and Automatic Relevance Determination. Finally, contains a fully functional package for real- and complex- valued maskable layers.