Changelog

Unreleased

v0.3.0 (2026-10-03)

Release note: v0.3.0

• Require Python 3.11 or newer and update the PyTorch, torchvision, Hugging Face Hub, and Pillow requirements.
• Add ConvNeXt, MobileOne, RepViT, character classifiers, and CTC text recognition, with expanded classification checkpoints and reference training tools.
• Add Adan and AdEMAMix, gradient accumulation, dataset analysis, and ONNX export and latency utilities.
• Secure remote checkpoint loading and restore UNet-family segmentation training and RepViT deployment conversion correctness.
• Repair YOLOv4 target assignment, BCE/CIoU losses, prediction initialization, AMP box geometry, and custom-anchor checkpoint loading.
• Correct shared confidence filtering, class-aware NMS, and detection evaluation matching while preserving YOLOv1/v2 loss behavior.
• Add a reproducible YOLOv4 CPU learning diagnostic with a frozen pretrained backbone, finite-gradient checks, normal-inference acceptance, and a saved checkpoint. Full CUDA/VOC training of the repaired implementation remains pending; YOLOv3 remains unimplemented as a detector.
• Expose detection LR finder bounds and progress, plot the complete recorded range, and make CodeCarbon quiet by default with --verbose-codecarbon available.

v0.2.1 (2022-07-16)

Release note: v0.2.1

v0.2.0 (2022-02-05)

Release note: v0.2.0

v0.1.3 (2020-10-27)

Release note: v0.1.3

v0.1.2 (2020-06-21)

Release note: v0.1.2

v0.1.1 (2020-05-12)

Release note: v0.1.1

v0.1.0 (2020-05-11)

Release note: v0.1.0