Skip to content

Latest commit

 

History

History
71 lines (54 loc) · 4.27 KB

CHANGELOG.md

File metadata and controls

71 lines (54 loc) · 4.27 KB
Raw

Changelog

Version 2022.06

  • COSMIT: accepted Black a the code style of choice, introduced pre-commit hooks for developers
  • FIX: having a dunder-version in the root of the package is a the standard (issue #24)
  • FIX: set the minimal python to 3.7 as pointed out in issue #24
  • UPD: bumped the base version of torch to at least 1.8
  • FIX: upgraded .utils.spectrum to new native torch complex backend (torch>=1.8)
  • FIX: ensured ONNX support in PR #14
  • ENH: implemented modulus-based maxpooling, requested in issue #17
  • FIX: made .Cplx instances deepcopy-able, fixing issue #18
  • DOC: improved docs for .nn.ModReLU indicating the sign-deviation from the original paper proposing it (issue #22)
  • DOC: added a basic TOC to the main README docs

Completed DEPRECATION cycles

  • misnamed VD and misplaced ARD layers in .nn.relevance
  • sparsity stats badly placed in .utils.stats
  • misnamed $\ell_0$ probabilistic pruning layer in .nn.relevance.extensions.real, since it had nothing to do with the Automatic Relevance Determination Bayesian approach

Version 2020.08.17

  • FIX: Fixed shape mismatch in .nn.init.cplx_trabelsi_independent_, which prevented it from working properly # 11
  • ENH: Hendrik Schröter implemented Complex Transposed Convolutions # 8, squeeze/unsqueeze methods for Cplx # 7, and added support for .view and .view_as methods for Cplx # 6
  • ENH: Introduce converters for special torch format of complex tensors (last dim is exactly 2) see torch.fft
  • ENH: Cplx now also has .size() method, which mimics torch.Tensor.size()
  • DOC: Improved documentation of .nn.casting modules

Version 2020.08

  • structure of the .nn.relevance was simplified
    • importing from nn.relevance.ard has been deprecated, and ARD layers have been moved to .real or .complex depending on their type
  • changed relevance layers class hierarchy in .relevance.real and .relevance.complex:
    • factored out Gaussian Local Reparameterization into pure *Gaussian layers, that reside in .real.base and .complex.base
    • subclassed Variational Dropout layers (*VD) from *Gaussian with improper prior KL mixin
    • subclassed ARD layers (*ARD) from Variational Dropout layers *VD with ARD Gaussian prior KL mixin

Version 2020.03

Major changes in .nn

  • The structure of the .nn sub-module now more closely resembles that of torch
    • .base : CplxToCplx and parameter type CplxParameter
    • .casting : real-Cplx tensor conversion layers
    • .linear, .conv, .activation : essential layers and activations
    • .container : sequential container which explicitly checks types of internal layers
    • .extra : 1-dim Bernoulli Dropout for complex-valued tensors (Cplx)
  • CplxToCplx can now promote torch's univariate functions to split-complex activations, e.g. use CplxToCplx[AvgPoool1d] instead of CplxAvgPool1d
  • Niche complex-valued containers were removed, dropped dedicated activations, like CplxLog and CplxExp

Major changes in .nn.relevance

  • misnamed Bayesian layers in .nn.relevance were moved around and corrected
    • layers in .real and .complex were renamed to Var Dropout, with deprecation warnings for old names
    • .ard implements the Bayesian Dropout methods with Automatic Relevance Determination priors
  • .extensions submodule contains relaxations, approximations, and related but non-Bayesian layers
    • \ell_0 stochastic regularization layer was moved to .real
    • Lasso was kept to illustrate extensibility, but similarly moved to .real
    • Variational Dropout approximations and speeds ups were moved to .complex

Enhancements

  • CplxParameter now supports real-to-complex promotion during .load_state_dict
  • added submodule-specific README's, explaining typical use cases and peculiarities

Prior to 2020.03

Prior version used different version numbering and although the layers are backwards compatible, their location within the library was much different.