Skip to content

Changelog

All notable changes to torchloop are documented here.
Format follows Keep a Changelog.
Versioning follows Semantic Versioning.

[0.3.0] - 2026-03-31

Added

  • Callback system with abstract Callback base class
  • WandBLogger callback for Weights & Biases integration
  • MLflowLogger callback for MLflow experiment tracking
  • callbacks optional dependency group: pip install torchloop[logging]
  • Edge deployment utilities (torchloop.edge)
  • FLOPs and parameter estimation (torchloop.edge.estimate)
  • MkDocs documentation site

Changed

  • Trainer now accepts callbacks: list[Callback] parameter
  • Trainer.fit() triggers on_train_begin, on_epoch_end, on_train_end hooks

[0.2.0] - 2026-03-28

Added

  • LR scheduler support in Trainer — any torch.optim.lr_scheduler
  • ReduceLROnPlateau handled automatically (passes val_loss)
  • Automatic Mixed Precision (AMP) via amp=True flag — CUDA only
  • LR logged per epoch in history["lr"]

Changed

  • Trainer.__init__ now accepts scheduler and amp parameters
  • Training log now includes current LR per epoch

[0.1.0] - 2026-03-27

Added

  • Trainer — PyTorch training loop with early stopping and checkpointing
  • Evaluator — classification report, confusion matrix, per-class F1
  • Exporter — PyTorch → ONNX → TFLite export pipeline
  • CI via GitHub Actions across Python 3.9, 3.10, 3.11
  • PyPI trusted publishing via OIDC