LibreSwin third-party attribution
=================================

Architecture reference
----------------------

The shared Swin V1 tower in ``nn.py`` is a native PyTorch implementation
derived from timm's Apache-2.0 Swin implementation:

    Repository: https://github.com/huggingface/pytorch-image-models
    Version:    v1.0.28
    Commit:     8ef73809f622e0031bd7f4940265734aef8b9978
    File:       timm/models/swin_transformer.py
    License:    Apache License 2.0
    Copyright:  Ross Wightman and the timm contributors

The classifier wrapper, family integration, conversion code, preprocessing,
postprocessing, and export tests are original LibreYOLO work. Parameter and
module names intentionally match timm so released tensors load strictly and
all four shipped variants produce bit-exact logits.

Model and weight lineage
------------------------

Swin Transformer V1 and its released checkpoints are from Microsoft Research:

    Repository: https://github.com/microsoft/Swin-Transformer
    Commit:     f82860bfb5225915aca09c3227159ee9e1df874d
    Storage:    https://github.com/SwinTransformer/storage
    Release:    v1.0.0
    Commit:     3cc359915d3a6079b176a871f68d5fb0d8dfdea2
    License:    MIT License

LibreYOLO converts the Tiny, Small, Base, and Large patch-4/window-7 classifiers
mirrored by timm on Hugging Face. Tiny, Small, and Base are ImageNet-1k
checkpoints. Large was pretrained on ImageNet-22k and fine-tuned on ImageNet-1k.
Learned tensors are unchanged; conversion adds LibreYOLO metadata and canonical
ImageNet-1k class names. Every hosted checkpoint repository carries the MIT
license and this provenance.

License texts
-------------

The Apache License 2.0 text is bundled at ``licenses/Apache-2.0.txt``. The MIT
license for each converted weight is shipped verbatim in its Hugging Face
repository.
