Changelog

All notable changes to FisherAdapTune will be documented here.

The format follows Keep a Changelog.

[Unreleased]

FisherAdapTuneTrainer: model-agnostic plug-and-play trainer with Fisher-guided iterative chunk freezing
AdaFisher: diagonal Fisher Information Matrix optimizer used internally for Fisher statistic collection
fisher_core: chunking, Jensen-Shannon divergence tracking, masking, and iterative freeze logic
EarlyStopping, save_checkpoint, plot_js_history utilities
SAM2 crack segmentation application (crack_segmentation/) with YAML config and full CLI
Minimal runnable example using synthetic data (examples/minimal_image_classifier.py)
Support for Linear, Conv2d, BatchNorm2d, and LayerNorm layers
Decoupled weight decay applied only to active (unfrozen) parameter entries
Optional W&B logging and matplotlib JS-distance grid plots
pyproject.toml packaging, Makefile automation, pre-commit hooks