LibreViT - third-party attribution
==================================

The native Vision Transformer architecture in ``nn.py`` is derived from:

    Source:  https://github.com/huggingface/pytorch-image-models
    File:    timm/models/vision_transformer.py and supporting timm/layers modules
    Version: v1.0.28
    Commit:  8ef73809f622e0031bd7f4940265734aef8b9978
    Author:  Ross Wightman and the timm contributors
    License: Apache License 2.0

LibreYOLO retains only the classic fixed-224 patch-16 classifier graph used by
the shipped AugReg ImageNet-1k checkpoints. Dynamic image sizes, alternate
pooling, register tokens, patch dropout, layer scale, stochastic depth, and
other later extensions are omitted. Module and parameter names remain aligned
with timm so all learned tensors load unchanged and exact inference parity can
be tested.

The architecture and AugReg checkpoint lineage originate from:

    Source:  https://github.com/google-research/vision_transformer
    Commit:  64801f1b3b367b3611cc27a3d45cc22870a36fb3
    Authors: Alexey Dosovitskiy et al., Google Research
    License: Apache License 2.0

The released checkpoints are the Apache-2.0 timm conversions of Google's
AugReg ImageNet-21k pretraining followed by ImageNet-1k fine-tuning. Each
published LibreYOLO weight repository carries the upstream Apache-2.0 license
and its own attribution notice.

Apache License 2.0: https://www.apache.org/licenses/LICENSE-2.0
