Skip to content

Latest commit

 

History

History
106 lines (88 loc) · 5.49 KB

VERSION.md

File metadata and controls

106 lines (88 loc) · 5.49 KB
Raw

iNNvestigate Changelog

Version 2.1.2

  • Maintenance Support TensorFlow 2.14 (#322)

Version 2.1.1

  • Maintenance Support TensorFlow 2.13 (a5395f8)
  • Maintenance Fix matplotlib deprecation of plt.cm (488a0fe)

Version 2.1.0

  • Feature Add way to strip arbitrary output activations (#310)
  • Bugfix Fix is_layer_at_idx for LRP (#308)
  • Maintenance Support Python 3.11 and TensorFlow 2.12 (#313)
  • Maintenance Update CI workflows and dev-dependencies (#313)

Version 2.0.2

  • Bugfix Fix Perturbate on RGB images (#306)
  • Bugfix Fix documentation
  • Maintenance Support TensorFlow 2.11 (#300)
  • Maintenance Update CI workflows and dev-dependencies (#301)

Version 2.0.1

  • Maintenance Remove dead analyzer.fit code left from PatternNet and PatternAttribution (#289)
  • Documentation Fixes to README and documentation

Version 2.0.0

iNNvestigate for TensorFlow 2. This is a major version release and therefore breaking backward compatibility.

Breaking changes:

  • update lower dependency bounds to Python 3.8 and TensorFlow 2.6
  • use TensorFlow's Keras instead of deprecated stand-alone Keras
  • manual disabling of eager execution is required via tf.compat.v1.disable_eager_execution() (#277)
  • temporarily remove PatternNet, PatternAttribution, LRPZIgnoreBias and LRPEpsilonIgnoreBias (#277)
  • remove DeepLIFT (#257)

Changes for developers:

  • switch setup to Poetry (#257)
  • adopt src and tests layout (#257)
  • adopt Black code style (#247)
  • add linters to dev dependencies (#257)
  • added type annotations (#263, #266, #277)
  • added reference tests & CI to guarantee identical attributions compared to v1.0.9 (#258, #277)
  • refactor backend (#263, #277)
  • refactor analyzers: explicit class attributes, fixes for serialization (#266, #277)
  • bug fixes (#263)

Version 1.0.9

  • BatchNormalization Layer compatible with LRP
  • EmbeddingLayer support
  • new Alpha-Beta-LRP-rules

Additionally various PR were merged and bugs fixed, for details see PR #222

Version 1.0.8

Bugfixes, increased code coverage, CI.

Version 1.0.7

Add Python 2 compatibility again.

Bugfixes.

Version 1.0.6

  • Add beta version of DeepLIFT (as in Ancona et.al.) and wrapper for DeepLIFT package.
  • Updating readme and bugfixes.

Version 1.0.5

Treat IntegratedGradients as attribution method and bugfixes.

Version 1.0.1-1.0.4

Added the following functionality:

  • Additional notebooks.
  • Analyzers: Input*Gradient
  • Added parameter to choose between plain, abs, square gradient in Gradient analyzer.
  • New interface via register-methods in analyzer base code.
  • Many fixes.
  • Support for read-the-docs documentation.

Version 1.0.0

Includes the following functionality:

  • Analyzers: Gradient, SmoothGrad, IntegratedGradients, PatternNet, PatternAttribution, LRP, DeepTaylor, Input, Random.
  • Pattern computer.