Skip to content

CHANGELOG.md

Latest commit

 

History

History
85 lines (68 loc) · 4.42 KB

CHANGELOG.md

File metadata and controls

85 lines (68 loc) · 4.42 KB

Changelog for the evolutionary package

0.10.0 (30.03.2026)

  • BREAKING: Upgrade dependency baselines across extras to current compatible releases.
  • Add explicit tensorboard dependency for prompt embedding visualization support.
  • Validate full .[all] installation and imports on Python 3.12 and Python 3.13.
  • Update notebooks for compatibility with the upgraded package structure and Colab install command.
  • Accept both colorfulness and colorfullness in CLIP-IQA evaluator, normalizing to TorchMetrics prompt keys.

0.9.0 (01.10.2025)

  • Added NSGA-III algorithm for many-objective optimization with different selectors

0.8.2 (09.08.2025)

  • Added image-to-prompt functionality in evolutionary_imaging.img2prompt that generates Stable Diffusion prompts from images using Qwen2.5-VL-7B-Instruct
    • Added a pre-testing notebook img2prompt_inference.ipynb to test this functionality
  • Added a Single-Objective SumEvaluator to sum the fitness values of multiple evaluators

0.8.1 (03.08.2025)

  • Fix completed_generations compatability with visualizations by increasing it in algorithm_base.py instead of in the individual algorithms.
  • Avoid recalculating best solution in NSGA_II by caching it
  • Added the AestheticPredictorV25ImageEvaluator

0.8.0 (07.07.2025)

  • Official support for sound generation using AudioLDM and evaluation with evolutionary_sound (sound subpackage) nad Audiobox Aesthetics.
  • BREAKING: prompt_embedding now supports sound generation but will not include imaging anymore since sound does not use it. You will now have to install both evolutionary[prompt_embedding] and evolutionary[sound or imaging].
  • Updating notebooks to reflect the package changes.

0.7.0 (16.03.2025)

  • Added TensorboardEmbedVisualizer to allow for easy visualization of embeddings with different algorithms like PCA, UMAP, etc. TensorBoard comes with a lot of functionality for that.
  • load_torch_model weights only is now True
  • plotting.py now supports post_evaluation time tracking
  • History of solutions (parents, mutations) now tracked in statistics.py with dataclasses in history.py
  • Family tree visualization made possible with graphviz, updated image processing.py and added family_tree.py

0.6.0 (18.01.2024)

  • Update transformers, diffusers and datasets to fix dependency issues with collab

0.4.3 (28.07.2024)

  • Fix numpy version to 1.26.4 to avoid 2.0 incompatibility

0.4.2 (07.04.2024)

  • added completed_generations to Algorithm to allow for better access to generations when run was not finished.

0.4.1 (01.04.2024)

  • Strict OSGA can be enabled in GA
  • ArithmeticCrossover now supports proportion, a parameter to control the amount of crossover
  • Statistics time tracking is now handeled by Algorithm implementations. IslandModel uses sum of all islands for time statistics.

0.4.0 (27.03.2024)

  • Breaking changes to Crossover/Mutation
    • moved tensor_variation to model_helpers.
    • Mutation does not modify the original tensor anymore
    • changed parameter names (crossover_rate, to interpolation_weight to avoid confusion)
  • GA, NSGA-II now use crossover- and mutation-rate
  • Breaking changes to Fitness-Statistics (avg, best, worst), now own class in statistics.py
  • Can now plot time spent on evaluation and creation of individuals

0.3.0 (25.03.2024)

  • Added GoalDiminishingEvaluator, CappedEvaluator
  • Added 'ring' and 'random' topology to the IslandModel to allow control of migration
  • Group-By-Ident visualization for image grid to allow visualization of best image per island
  • Radar chart visualization for multi-objective optimization now supports max-value, so the charts remain the same size
  • Fixed RankSelector division by zero error

0.2.6 (22.03.2024)

  • NSGA-II can now use a binary tournament selection for the selection of parents

0.2.5 (20.03.2024)

  • NSGA-II now returns result which has the best fitness sum with optional normalization

0.2.4 (08.03.2024)

  • Add optional identifiers to image-saving do differentiate between islands in the island modal

0.2.3 (08.03.2024)

  • Fixed plotting when only one variation (ex. avg_fitness) is plotted
  • Added generation argument to Algorithm.perform_generation
  • NSGA now supports callback after NDS

0.2.2 (06.03.2024)

  • Fixed imports, structure to work with colab

0.2.0 (06.03.2024)

  • Refactored package structure to single setup.py with extras, which allow a finer selection of dependencies and packages.