Skip to content

Latest commit

 

History

History
55 lines (46 loc) · 3.05 KB

CHANGELOG.md

File metadata and controls

55 lines (46 loc) · 3.05 KB
Raw

Changelog for deepstruct

0.10

  • bug-fix for 'RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation' caused by in-place operation in model (thanks to Mohammad Alahmad)
  • new methods for scalable models from a underlying computational theme structure
  • more explicit methods to store and re-load a cached layered graph
  • some documentation on base module MaskedLinearLayer
  • dependency updates

0.9

  • re-introduced saliency as an optional additional property on MaskedLinearLayers for communicating saliency measures on weight-level to decide on further pruning
  • fixed some of the simpler pruning functions such as prune_network_by_saliency() and prune_layer_by_saliency() from deepstruct.pruning
  • masks up to now do not consider bias vectors which might be unexpected behaviour

0.8

  • deprecation of learning utilities
  • integrated additional normalization layers
  • masks on maskable layers are parameterizable to investigate on structural regularization ideas
  • functional dataset can now be easily stored in a pickle file

0.7

  • new minimal version requirement is python 3.7
  • introduced interface for "functors" which transform a nn.Module into a directed acyclic graph
  • created a first functor for Linear and MaskedLinear layers
  • a graph transform class passes a random input through a generic module and can transform it into a graph given that it consists of linear or conv2d layers (first tests added)
  • added mkdocs to provide an initial documentation skeleton

0.6

  • introduced BaseRecurrentLayer, MaskedRecurrentLayer, MaskedGRULayer, MaskedLSTMLayer
  • introduced deepstruct.recurrent.MaskedDeepRNN for sparse recurrent models

0.5

  • new feature: concept of scalable families which is a first notion of graph themes analysis
  • file restructuring for better semantics
  • pypaddle will be renamed to deepstruct

0.4

  • switched to poetry for dependency and build management
  • added integration tests
  • switched to pytest instead of unittest

0.3

  • added support to define input shape for MaskedDeepFFN and MaskedDeepDAN
  • changed parameter for recompute_mask(epsilon) to recompute_mask(theta) as it should denote a threshold
  • implemented a first running version of a randomly wired cell network, more general than RandWireNN and in spirit of analysing graph theoretic properties
  • bugfixes on generating structures from masks
  • added/modified data loader utilities for mnist/cifar (probably no official part and concern of this library tools)
  • fixed PyPi setup and tested installation routine
  • defined networkx and torch as dependencies in setup.py. Next will be to check if it can be shadowed by pytorch packages from conda channels
  • added a DeepCellDAN() which builds directed, acyclic networks with customized cells given a certain structure

0.2

  • introduced LayeredGraph as a wrapper for directed graphs which provides access to its layered ordering
  • central provided modules are MaskedLinearLayer, MaskedDeepFFN and MaskedDeepDAN
  • provided first functionality to generate structures from masked modules